It is Monday morning.

You walk into your office wearing what appears to be an ordinary pair of glasses. They look no different from the ones millions of people already wear every day. They help you navigate, translate foreign languages, answer questions, remember where you parked your car and identify the restaurant you passed ten minutes ago.

They also quietly observe almost everything you do.

They notice who you meet, where you walk, what you look at, what you read, who speaks to you, how long conversations last and what decisions are made. Every meeting becomes searchable. Every conversation becomes queryable. Every day becomes a structured dataset.

Now imagine arriving at work.

Your calendar assistant has already summarised your first meeting. Your AI meeting companion knows the history of every discussion with that client over the last three years. It reminds you that the CFO was unconvinced by a particular proposal eighteen months ago, that Legal raised a concern during contract negotiations in February, and that Engineering quietly abandoned a similar idea two years earlier.

None of this feels particularly remarkable.

In fact, it feels useful, very useful, and that is precisely why it is happening.

For years, we assumed artificial intelligence would become smarter by reading more books, more web pages and more documents.

That phase is largely over.

The world’s largest models have already consumed an extraordinary proportion of publicly available text. Their next challenge is no longer acquiring more information.

It is acquiring more context:

Information tells AI what happened.

Context teaches AI why it happened.

That is why almost every major technology company is now investing in systems that observe how humans actually work.

The objective is no longer simply to answer questions, it is to understand judgement.

Orwell saw surveillance. He missed the business model.

George Orwell’s Nineteen Eighty-Four remains one of the most powerful warnings ever written about surveillance. His world was one in which citizens were continuously watched by the state. Surveillance existed to enforce obedience. Information flowed upwards towards those in power.

The modern version looks remarkably different:

• Nobody forces millions of people to buy smart glasses.

• Nobody insists they install AI meeting assistants.

• Nobody mandates AI note takers, workflow analysers or digital companions.

People adopt them because they are genuinely useful:

• They save time.

• They remember forgotten details.

• They draft emails.

• They summarise meetings.

• They identify action items.

• They retrieve information that would otherwise disappear into forgotten notebooks or archived emails.

The commercial incentive is productivity.

The technical consequence is something much larger. Every one of those interactions teaches the machine a little more about how humans solve problems:

Orwell imagined surveillance as political infrastructure.

The AI era is creating surveillance as learning infrastructure.

That distinction changes almost everything.

Why AI suddenly wants to watch you

One of the biggest misconceptions about artificial intelligence is that knowledge is the scarce resource. It isn’t.

The scarce resources are judgement and willingness to act.

Ask an AI how to calculate discounted cash flow and it will probably produce an excellent explanation.

Ask it how to persuade a reluctant client, negotiate a difficult supplier contract, recognise that a project is quietly failing despite apparently healthy metrics or identify the moment when an experienced engineer becomes uncomfortable with a proposed design.

Those are much harder problems, not because they require more intelligence, but because they require experience.

Humans rarely write down how they make decisions.

Policies explain what organisations intend to do.

Meetings reveal what they actually do.

Company handbooks describe escalation procedures.

Experienced managers know when to ignore them.

Standard operating procedures explain the rules.

Culture determines which rules matter.

The most valuable organisational knowledge has always existed between the lines.

For decades that knowledge simply disappeared:

• People retired.

• Teams reorganised.

• Lessons were forgotten.

• Mistakes were repeated.

Now something fundamentally different is happening.

Not because somebody explicitly documents them, because millions of conversations quietly contain them.

The venture capital firm Andreessen Horowitz recently argued that AI needs to be onboarded much like a new employee. You do not ask a new recruit simply to read the company handbook. You invite them into meetings because meetings contain the unwritten culture of an organisation.

That observation is far more profound than it first appears.

Meetings are where judgement lives. They contain uncertainty, disagreement, compromise, political realities, customer nuance, risk appetite and countless exceptions that never make their way into official documentation.

If AI can observe enough of those conversations, it begins to develop something far more valuable than knowledge, it develops organisational context. And organisational context is rapidly becoming one of the most valuable assets in the AI economy.

Behaviour has become the new intellectual property

For years, executives were told that “data is the new oil.”

It was an appealing phrase.

It was also incomplete.

Raw data has become abundant.

Most organisations possess more operational data than they know how to exploit. Public information has become increasingly commoditised. Foundation models have already ingested enormous quantities of publicly available content.

Behaviour is different.

Every organisation behaves differently:

Two banks may use identical software. Their credit committees will not make identical decisions.

Two insurers may have identical underwriting manuals. Their experienced underwriters will reach different conclusions.

Two consulting firms may use the same methodologies. Their partners will solve client problems differently.

Those differences are difficult to document but they are exceptionally valuable to observe.

AI companies increasingly understand that behaviour cannot simply be scraped from the internet, it must be witnessed.

That is why the race has quietly shifted from collecting documents to collecting context:

• Every recorded meeting.

• Every customer interaction.

• Every software commit.

• Every workflow.

• Every correction.

• Every exception.

• Every approval.

• Every hesitation.

Every one of them teaches the system something that no policy manual ever could.

The rise of the Memory Economy

Perhaps the most important shift is that organisations are no longer merely building databases, they are building memory:

• Traditional enterprise software stored transactions.

• CRMs remembered customers.

• ERPs remembered inventory.

• Document management systems remembered files.

• The next generation remembers conversations.

Not simply that a meeting occurred:

• What was discussed.

• Who objected.

• Which assumptions changed.

• Which risks were dismissed.

• Which promises were made.

• Which decisions were later reversed.

This is not simply better record keeping, it is a completely different form of organisational infrastructure.

Human organisations have historically suffered from institutional amnesia and knowledge walked out of the building every Friday evening: sometimes permanently.

Every resignation, retirement or acquisition meant rebuilding context that had taken years to accumulate. Persistent organisational memory changes that equation.

The company itself begins to remember and that has enormous benefits:

• It also raises entirely new questions.

• Who owns that memory?

• Who may search it?

• Who may train on it?

• And what happens when the organisation remembers more accurately than the people inside it?

The Boardroom point

Every significant technological shift eventually changes what boards are expected to govern.

The Industrial Revolution required boards to understand factories. The Information Age required them to understand software. The AI era requires them to understand something far less tangible: organisational memory.

Until recently, memory was largely a human characteristic. It resided in experienced employees, long-serving managers and trusted advisers. Boards worried about key-person dependency because they understood that knowledge disappeared when people did.

Artificial intelligence changes that assumption: the organisation itself is beginning to remember.

Every recorded meeting, customer call, engineering review, underwriting discussion, investment committee debate and strategic planning session contributes another layer of institutional knowledge. Individually, none of these interactions appears particularly significant. Collectively, they become a searchable representation of how the organisation actually thinks.

That distinction matters because companies have always had two operating systems.

1. The first is formal. It consists of organisational charts, policy manuals, procedures, delegated authorities and governance frameworks.

2. The second is informal. It consists of judgement, relationships, trust, unwritten conventions and decades of accumulated experience.

The first can be documented, the second has historically been impossible to capture, until now.

Boards therefore face a new fiduciary question:

The answer is unlikely to be “none”. The productivity gains are simply too compelling. Better onboarding, faster decision-making, improved continuity, fewer repeated mistakes and stronger organisational resilience all argue in favour of richer institutional memory.

But neither is the answer “everything”, because the moment judgement becomes data, it becomes an asset.

Like every strategic asset, it requires governance.

There is no such thing as free context

Most organisations currently evaluate AI using familiar procurement questions:

• How much will it cost?

• How much time will it save?

• Which departments should adopt it first?

Those remain sensible questions but they are no longer sufficient:

• Every AI deployment also creates a less visible economic exchange.

• Employees receive assistance.

• The model receives experience.

That exchange is remarkably easy to overlook because it feels so natural.

• An AI meeting assistant produces excellent notes.

• A coding assistant accelerates software development.

• A customer service copilot recommends responses.

• An AI sales coach analyses successful conversations.

• Each application genuinely improves productivity.

Each application also observes how skilled professionals perform their work and this is not necessarily sinister. It is simply how modern machine learning improves. Large language models already possess broad knowledge of the world.

What they increasingly lack is deep understanding of how particular organisations operate. That understanding can only be learned from observing real work.

The CFO version

Finance leaders have always excelled at valuing tangible assets:

Cash.

Property.

Inventory.

Receivables.

They have also become increasingly sophisticated in understanding intangible assets such as brands, software and intellectual property.

Behavioural capital represents something different again, it rarely appears anywhere on the balance sheet. Yet it may become one of the most valuable assets an organisation possesses.

Consider two businesses with identical technology platforms, identical customer bases and identical products.

One consistently outperforms the other.

Why?

The answer rarely lies in the software, it lies in thousands of accumulated decisions made by experienced people over many years:

• How pricing exceptions are handled.

• Which customers receive flexibility.

• When risks are accepted.

• When projects are terminated.

• When investment accelerates.

• When management intervenes.

• These decisions create economic value.

Historically they remained trapped inside human experience, today they are becoming observable. This creates a strategic question that finance leaders have rarely needed to ask before.

If behavioural knowledge becomes trainable, who captures the value?

The enterprise?

The technology provider?

The cloud platform?

The model developer?

Or some combination of all four?

Current accounting standards have little to say about behavioural capital because they evolved in a world where behaviour could not easily become an asset, AI changes that assumption.

The next generation of balance sheets may never explicitly recognise organisational memory.

Investors almost certainly will.

The CEO version

Chief executives often describe culture as the organisation’s greatest competitive advantage, unfortunately, culture is also one of its least measurable.

Ask five executives why their company consistently wins difficult contracts and you will probably receive five different answers:

• Some will point towards customer relationships.

• Others towards engineering excellence.

• Others towards operational discipline.

• All may be correct.

• None fully explains why the organisation behaves differently from its competitors.

Behaviour emerges from thousands of invisible decisions:

It is how difficult customers are handled.

How mistakes are admitted.

How junior employees challenge senior colleagues.

How risk committees balance growth against caution.

How people behave when no policy explicitly tells them what to do.

For decades, CEOs accepted that much of this knowledge simply could not be documented, it was spread through observation:

• New employees learned by sitting beside experienced colleagues.

• Graduate trainees absorbed judgement through repeated exposure to real situations.

• Managers developed intuition by participating in hundreds of meetings.

The process was slow, expensive and entirely human.

Artificial intelligence introduces another possibility:

Instead of asking new employees to absorb years of organisational behaviour, organisations can increasingly expose AI to those same interactions.

The machine begins developing organisational context alongside the employee.

That offers extraordinary opportunities:

Imagine reducing executive onboarding from twelve months to three.

Imagine preserving the accumulated experience of retiring specialists.

Imagine integrating acquisitions without losing decades of institutional knowledge.

Imagine querying twenty years of strategic discussions before making today’s decision.

These capabilities move beyond productivity, they begin reshaping organisational resilience itself.

Yet they also introduce a profound strategic risk:

Every organisation believes its competitive advantage lies in products, patents or capital.

Increasingly, it may lie elsewhere.

It may lie in judgement.

The question is uncomfortable, but it deserves asking.

Share Something For The Weekend

The Board version

Board discussions around artificial intelligence frequently begin with regulation. They move quickly towards cybersecurity, privacy, bias, explainability and legal compliance. Those topics remain critically important.

But another deserves equal attention:

Institutional sovereignty.

Every organisation must decide where its collective memory resides.

Inside the company.

Inside its own private infrastructure.

Or increasingly, inside commercial AI ecosystems operated by third parties.

This is not simply an information security question, it is a governance question:

• What happens when every board paper, every strategy session, every acquisition discussion, every underwriting committee and every product review contributes another layer of machine-readable institutional knowledge?

• Who controls access?

• Who determines retention?

• Who decides whether that knowledge contributes to future model improvement?

The answers differ across jurisdictions, technologies and contractual arrangements.

But one principle remains remarkably consistent:

If organisations fail to govern their behavioural capital, somebody else eventually will.

Boards have spent decades protecting financial capital, the next decade may require protecting organisational memory with equal care.

The hard truth

Much of the public debate around artificial intelligence assumes the technology is primarily replacing labour but that is too narrow. Artificial intelligence is first learning labour, only afterwards can parts of that labour be automated.

Every AI system that drafts documents, analyses contracts, recommends underwriting decisions or assists software developers became capable by observing enormous quantities of human work, and that process has not stopped.

It is accelerating:

The greatest AI companies may not ultimately possess the smartest algorithms, they may simply possess the richest behavioural datasets.

The scarce resource would no longer be computational power alone, nor access to public information. It would be privileged visibility into how experienced people actually think.

Perhaps Orwell’s greatest mistake was not imagining surveillance, it was assuming surveillance existed primarily for control.

The AI economy suggests another possibility:

The machine is not watching because it wants obedience.

The machine is watching because it wants an education.

The weekend takeaway

For centuries, organisations have accumulated knowledge one conversation at a time. Most of those conversations disappeared almost as quickly as they occurred. Human forgetting was not a flaw, it was simply how organisations functioned.

Artificial intelligence is quietly ending that era.

For the first time in history, companies can build persistent organisational memory that grows richer with every meeting, every project and every decision. That capability promises extraordinary improvements in productivity, continuity and resilience.

It also creates a new strategic reality.

The organisations that succeed in the AI era may not simply be those that deploy the best models.

They may be those that most carefully govern what those models are allowed to learn.

Because the defining question of 2026 is no longer whether Big Brother is watching:

？Questions to think about this weekend

In this new section, I leave you with a few questions to ‘percolate’:

• If AI assistants remember everything, what should organisations deliberately choose to forget?

• When does organisational memory become intellectual property?

• Should employees help train systems that may eventually automate parts of their work?

• If two companies buy exactly the same AI model, does competitive advantage increasingly come from the behavioural context they provide?

• Has “knowledge work” quietly become “behaviour work”?

  Leave a comment

🧠 Techie Corner for the Non-Techies

One of the biggest misconceptions about AI is that it “knows everything.” It doesn’t.

Today’s most capable Large Language Models (LLMs) possess enormous general knowledge because they have been trained on large quantities of public text. However, they know very little about your company. They do not understand your customers, your suppliers, your unwritten conventions or how your board really makes decisions.

That is why the AI industry has shifted its attention from collecting more information to collecting more context.

Here are some of the concepts that increasingly matter:

1. Foundation Model

A foundation model is a very large AI model trained on vast amounts of public information before it is ever deployed inside an organisation. Think of it as a highly educated graduate. It knows history, science, finance, programming and languages.

What it does not know is how your business works.

That knowledge has to be added later.

2. Context

Context is everything surrounding a decision. The same sentence means different things depending on who said it, why they said it, what happened beforehand and what everyone already knows. Humans constantly use context without noticing.

AI does not.

That is why context has become one of the most valuable resources in enterprise AI. The model may understand insurance but it does not automatically understand your underwriting philosophy.

3. Context Window

A context window is the amount of information an AI model can consider at one time. Early systems could only remember a few pages of text. Modern frontier models can analyse entire reports, books or software repositories during a single conversation.

Larger context windows make richer reasoning possible, they also increase computing cost. So, more context is not automatically better, relevant context is.

4. Long-Term Memory

Most early chatbots forgot everything once a conversation ended. In contrast, modern AI increasingly maintains persistent memory:

• It remembers preferences.

• Recurring projects.

• Frequently used terminology.

• Previous conversations.

• Important decisions.

In organisations this becomes even more valuable. Imagine an AI that remembers every client meeting over the past five years rather than only today’s conversation.

That changes the quality of assistance dramatically.

5. Retrieval-Augmented Generation (RAG)

One of the most important enterprise AI techniques is called Retrieval-Augmented Generation, usually shortened to RAG. Instead of relying only on what it learned during training, the AI first searches trusted internal information before generating an answer.

For example, before answering a question about an insurance programme it might retrieve:

• underwriting guidelines

• previous submissions

• policy wordings

• engineering reports

• claims histories

• board decisions

Only then does it formulate a response. RAG helps reduce hallucinations because the AI reasons over current company information rather than relying purely on memory.

6. Embeddings

Computers do not naturally understand meaning. Embeddings are mathematical representations that allow similar ideas to sit close together.

A document discussing “renewable energy insurance” might be placed close to another discussing “offshore wind risk”, even if neither uses identical words.

Embeddings allow AI to search by meaning rather than keywords. That makes organisational memory dramatically more useful.

7. Vector Databases

Traditional databases search for exact matches. Vector databases search for similar ideas.

Instead of asking:

“Find documents containing these exact words.”

they ask:

“What previous situations most resemble this one?”

That is much closer to how experienced executives think. An underwriter rarely asks whether a submission contains identical wording. They ask whether it resembles another risk they have seen before.

Vector databases make that possible for machines.

8. Multimodal AI

Early AI systems worked primarily with text. Modern models increasingly understand multiple forms of information simultaneously:

• Text.

• Images.

• Audio.

• Video.

• Computer code.

• Sensor readings.

• Eventually robotics.

Smart glasses illustrate this evolution:

• They are not simply recording video.

• They combine visual information, spoken conversations, location, time and movement into one continuously evolving stream of context.

9. Behavioural Data

Behavioural data differs fundamentally from traditional business data:

Traditional systems record what happened.

Behavioural systems record how it happened.

• Which document did someone open first?

• Which version did they reject?

• Who did they consult?

• Which alternative did they consider?

• Where did they hesitate?

These patterns often reveal judgement far better than final outcomes alone. That is why behavioural data has become strategically valuable.

10. Digital Memory

Human memory is imperfect:

• We forget names.

• We misremember dates.

• We compress events.

• We simplify complicated histories.

Digital memory behaves differently:

• It retrieves exactly what was recorded.

• Sometimes that is advantageous.

• Sometimes it preserves context humans would rather forget.

This creates an important governance question:

Should organisations optimise for perfect recall?

Or should they deliberately preserve the human ability to forget?

That question is becoming technological rather than philosophical.

11. Behavioural Cloning

Behavioural cloning is an AI technique in which a machine learns by observing examples of human behaviour. Instead of programming every rule explicitly, engineers allow the system to watch experts repeatedly perform a task. The AI gradually learns to imitate those patterns.

That approach has become increasingly important in robotics, autonomous vehicles and enterprise AI. It also explains why recorded work has become so valuable. Machines do not merely need instructions, they need demonstrations.

12. Model Context Protocol (MCP)

One of the newest developments in enterprise AI is the Model Context Protocol, usually abbreviated to MCP.

Think of MCP as a ‘universal connector’. Instead of building a separate integration for every application, MCP allows AI systems to communicate with many different tools through a common protocol.

For example, one AI assistant might simultaneously retrieve:

• today’s calendar

• CRM information

• engineering drawings

• SharePoint documents

• financial models

• regulatory guidance

without needing a bespoke connection for each one.

Many technologists believe MCP may become one of the foundational standards underpinning enterprise AI over the coming years.

None of these technologies are remarkable in isolation. Together they create something genuinely new.

That is a much bigger idea than simply building better chatbots.

📚 Further reading

• George Orwell: Nineteen Eighty-Four (1949)

  Published on 8 June 1949, Orwell's classic novel remains the defining exploration of surveillance, information control and the relationship between knowledge and power. The AI era differs fundamentally because today's surveillance is often driven by commercial incentives rather than state coercion, but Orwell's central insight (that information changes power) remains highly relevant.

• David Haber (Andreessen Horowitz): Everything is Recorded Now (2026)

  One of the clearest explanations of why AI increasingly requires organisational context rather than simply more documents. The article argues that meetings contain the tacit knowledge that company manuals never capture, and that AI must observe organisations much as new employees do.

• Hannah Murphy (Financial Times): Meta tests ‘super sensing’ AI glasses that can capture every moment (2026)

  An important article describing Meta’s reported work on continuously sensing AI glasses capable of capturing a wearer’s surroundings throughout the day. Beyond the technology itself, it raises difficult questions about privacy, consent, organisational memory and the emergence of always-on AI assistants.

• Dhruv Mehrotra & Dell Cameron (WIRED): Meta Silently Added Face-Recognition Code for Its Smart Glasses to Millions of Phones (2026)

  A technical investigation into software components reportedly embedded within Meta’s smart-glasses ecosystem. It illustrates how modern AI increasingly combines visual recognition, behavioural data and persistent identity rather than simply capturing photographs.

• Jared Spataro (Microsoft): The 2025 Annual Work Trend Index: The Frontier Firm is Born (2025)

  Microsoft’s annual Work Trend Index combines survey research with Microsoft 365 telemetry and LinkedIn labour-market analysis to examine how AI is changing organisations. As with any vendor-sponsored research, readers should bear in mind Microsoft’s commercial interests, but it remains one of the most comprehensive publicly available reports on enterprise AI adoption.

• Shoshana Zuboff: The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power (2019)

  Written before the emergence of generative AI, Zuboff’s book remains one of the most influential analyses of behavioural data as an economic resource. Her argument that companies increasingly compete by capturing and monetising human behaviour has become even more relevant as AI systems seek richer behavioural context rather than simply larger datasets.

• OpenAI: Introducing ChatGPT Agent (2025)

  OpenAI’s announcement of ChatGPT Agent illustrates the industry’s shift from conversational AI towards systems capable of executing multi-step tasks across applications. It provides useful insight into why organisational context, memory and tool access are becoming increasingly important components of enterprise AI.

• Anthropic: Introducing Claude for Enterprise (2024)

  Anthropic explains its approach to enterprise AI, including large context windows, retrieval from internal knowledge bases and governance controls. The material is particularly useful for understanding why organisational memory is becoming a strategic enterprise capability rather than simply another chatbot feature.

Last week on Something for the Weekend

Last week, we argued that the workflow was always the asset.

Enterprise AI creates value not because it writes better emails or generates more code, but because it has the potential to redesign how work moves through an organisation. The companies that succeed will measure verified business outcomes rather than token consumption, redesign workflows rather than automate old habits, and treat digital co-workers as managed members of the workforce rather than clever software.

The CFO receives the AI update before the board pack is finalised.

The dashboard looks impressive:

AI adoption is up.

Tokens processed are up.

Engineering is using coding agents.

Legal is testing contract review.

Procurement has started to use negotiation assistants.

Customer service is routing more queries through copilots.

Several business units now report “AI-assisted productivity” in their monthly performance reviews.

The CIO is pleased. The CEO is encouraged. The board sees momentum.

Then the CFO opens the cost report.

The numbers are less reassuring:

The company is spending more on AI-enabled SaaS, more on APIs, more on model access, more on cloud infrastructure, more on integration, more on data preparation, more on cyber controls, more on human verification and more on the internal teams needed to make the new tools work.

Some of the spending is useful.

Some of it is simply machine activity.

The CFO asks the first question:

“Which of this AI spend has reached the P&L?”

The room becomes more cautious.

The CIO explains that some teams are already changing model defaults, routing simpler work to cheaper models, using prompt caching, shortening context windows and reserving expensive frontier models for harder tasks. That makes sense. Coinbase recently became a useful public example of this pattern: higher token usage did not automatically mean higher AI spend once default models, routing, caching, context discipline and spend visibility were changed.

Then the COO asks the better question:

“Why are we adding AI to old workflows instead of redesigning the work?”

That question changes the meeting.

The company already owns the important assets:

• It owns the customer relationships.

• It owns the supplier relationships.

• It owns the workflows.

• It owns the exception history.

• It owns the operating data.

• It owns the decision rights.

• It owns the brand, trust, permissions, plants, systems, contracts, approvals and physical infrastructure through which work gets done.

The issue is not whether the company owns enough assets to make AI useful.

The issue is whether those assets are trapped inside old workflows built for humans, old software built for screens, old approval chains built for slower work and old metrics built to prove adoption rather than value.

The digital co-worker does not only create a cost problem, but it exposes an operating-model problem.

The Boardroom point

Large corporates already own the scarce ingredients of enterprise AI.

Not just data.

Data by itself are often inert. They can be incomplete, duplicated, badly permissioned, incorrectly labelled, stuck in different systems or divorced from the context that gives them meaning.

The real corporate asset is the operating system of the business:

• Data.

• Workflow.

• Client insight.

• Client relationships.

• Supplier insight.

• Supplier relationships.

• Exception history.

• Tacit judgement.

• Regulatory permissions.

• Brand trust.

• Physical assets.

The practical knowledge of what happens when the process map meets reality. AI does not create that operating system but it makes visible whether management understands it.

That is why the current AI debate is shifting:

The first phase of enterprise AI was about access.

The second phase was about adoption.

The third phase will be about redesigning work.

The danger is that companies measure the wrong thing at precisely the moment when the cost structure is becoming more visible.

Prompts sent are not value. Tokens consumed are not value. Agents launched are not value. AI-enabled SaaS renewals are not value. A larger AI budget is not transformation. A smaller AI budget is not necessarily discipline.

The useful metric is verified machine work.

Did the digital co-worker do something economically useful? Was it correct? Was it used? Did it remove an old step? Did it reduce rework? Did it shorten cycle time after review? Did it reduce risk? Did it improve customer outcome? Did it lower total cost after tokens, cloud, integration, verification, governance and failure are counted?

That is the boardroom question.

The AI market is now forcing that question because the pricing model is changing. Reuters reported this week that usage-based AI pricing is making corporate AI spending harder to predict even as token prices fall, and that businesses are increasingly turning to smaller, cheaper models and routing tools to reserve premium models for more complex work. Gartner has also warned that token-driven AI coding spend is becoming difficult to justify where organisations lack the maturity to measure cost against business impact.

This is not a temporary procurement annoyance.

It is the beginning of machine-work accounting.

Tokenmaxing is not transformation

Tokenmaxing is the corporate habit of treating visible AI consumption as evidence of progress. It was understandable at first.

Boards wanted to know whether employees were using the tools. CIOs needed proof that licences were not sitting idle. Business-unit leaders wanted to show experimentation. Software vendors wanted usage. Investors wanted adoption. Employees wanted speed.

So usage became the early proxy:

• More prompts.

• More tokens.

• More copilots.

• More agent runs.

• More AI-assisted documents.

• More lines of AI-assisted code.

The problem is that usage was never the same as productivity.

A company can increase token consumption while leaving the old workflow intact. It can generate more documents without making better decisions. It can produce more code while increasing review burden. It can answer more customer questions while increasing repeat contacts. It can summarise more meetings while holding the same number of meetings.

That is not transformation, it is a new layer.

The most dangerous version is not reckless AI use. It is controlled, approved, measured AI use that still measures the wrong thing.

A company can comply its way into waste. It can build dashboards, policies, approved tools, training modules and governance committees, then reward teams for activity that never becomes verified value.

The question should not be:

How much AI are we using?

The better question is:

Which old work disappeared because the digital co-worker now performs part of it safely, cheaply and verifiably?

If the old work remains, AI has not transformed the workflow, it has added a meter to it.

The workflow was always the asset

Most large corporates have an asset they do not properly value because it does not sit neatly on the balance sheet:

It is the workflow.

The workflow is not the process map. The process map is the clean version. The workflow is what actually happens:

It is how a customer exception gets resolved when the standard policy fails. It is how a supplier is treated differently because everyone knows they recover faster after a shock. It is how a claims handler spots that a loss is not really like the last one. It is how a maintenance engineer knows that a machine will fail after a particular vibration pattern even when the dashboard is still green.

It is how an underwriter reads silence in a submission. It is how a finance team knows which invoice discrepancy is harmless and which one signals fraud. It is how a buyer knows which supplier promise is credible. It is how a plant manager knows when a system can be stopped safely and when stopping it creates a larger problem.

That knowledge is rarely stored in one place.

It lives across enterprise systems, emails, spreadsheets, contracts, call notes, service records, supplier histories, customer conversations, claims files, maintenance logs, workarounds, local relationships and people’s heads.

That is why generic AI can disappoint inside large companies:

The model may be capable, but it does not automatically understand the business. It can summarise the policy without knowing how the policy behaves under stress. It can draft the contract without knowing which clause has caused disputes. It can read the supplier scorecard without knowing which supplier absorbs pressure quietly and which one fails loudly.

The scarce asset is not data alone, it is data joined to workflow, relationships and judgement.

This is important for the make-versus-buy question:

The SaaS vendor may own the interface. The corporate owns the operating reality. If the vendor brings scale, resilience, regulatory assurance, security, network effects, maintenance and deep workflow capability, buying still makes sense.

But if the vendor mainly wraps a generic interface around the corporate’s own data, customer relationships, supplier relationships and workflow history, the economics are changing.

AI-enabled development does not make software free. Maintenance, testing, security, audit, user support, integration and accountability still cost money. But AI does lower the cost of creating workflow-specific tools. That shifts the boundary.

The new question is not:

Can we build software?

The better question is:

Does the vendor own something scarce, or are we paying rent on our own operating knowledge?

Gartner has described a related pressure as “agentic arbitrage”, where AI agents complete tasks across systems and reduce the need for users to interact with traditional software interfaces. Gartner estimates that up to $234 billion of enterprise application software spend could be exposed to this dynamic by 2030, roughly 20% of enterprise application SaaS spending.

That does not mean SaaS disappears, what it means is weak SaaS has to justify itself.

The digital co-worker needs a job description

The phrase “digital co-worker” is useful only if it is taken seriously.

A human employee does not enter the company with unlimited authority. The company defines the role, reporting line, budget, access rights, approval limits, escalation process, review standard and accountability structure.

A digital co-worker needs the same discipline.

It needs a job description.

Not a marketing description. A real operating specification.

What work may it perform? Which data may it access? Which tools may it call? Which systems may it change? Which model tier may it use? How much may it spend? When must it escalate? Which outputs require verification? Which decisions require a receipt? When must it stop?

Without that, the digital co-worker becomes unmanaged machine labour.

It can generate useful work. It can also generate expensive activity, plausible error, hidden dependency and verification debt. It can operate continuously. It can create outputs faster than the company can absorb them. It can make the old bottleneck reappear as senior review.

The point is not to slow everything down.

A digital co-worker that classifies invoices inside a narrow rule set is different from a digital co-worker that drafts a board paper. A digital co-worker that retrieves policy wording is different from one that recommends declining a claim. A digital co-worker that summarises supplier contracts is different from one that changes payment terms.

Authority matters.

So does reversibility.

So does evidence.

The company should be able to reconstruct what the digital co-worker did, what it saw, which model ran, which tool was used, which data were accessed, what output was produced, who approved it and whether the result was used.

“The AI said so” is not an audit trail.

Routing is the new operating discipline

A large company would not assign every task to the most expensive partner in the firm, and it should not assign every machine task to the most expensive model.

The expensive model should be the escalation point, not the default worker.

Routine extraction, classification, comparison, drafting, reconciliation and summarisation may not need frontier intelligence. Some work can go to smaller models. Some can go to retrieval systems. Some can go to deterministic software. Some can run locally because the data are sensitive. Some should use a frontier model because the reasoning burden is high. Some should stop and escalate to a human.

Gartner’s recent comments on AI coding costs are directionally consistent with this. It recommends breaking work into smaller tasks, aligning model selection with task complexity and reserving frontier models for complex, high-value development work. Reuters reported a similar market shift, with companies using routing tools such as OpenRouter, Martian or proprietary internal gateways to assign tasks to cost-effective models and reserve premium models for complex work.

That sounds technical, yet it is not only technical, it is workforce planning for machine labour.

Routing should be based on:

• Cost.

• Risk.

• Confidentiality.

• Complexity.

• Reversibility.

• Required judgement.

• Verification burden.

• Auditability.

The objective is not to minimise tokens. A cheap answer that creates review work, litigation risk or customer damage is not cheap. An expensive answer that prevents a serious error may be good value.

The objective is verified output at the lowest total cost and acceptable risk.

That is a different management discipline.

The make-or-buy question comes back

The old software answer was simple enough.

Buy.

Software development was slow. Internal IT capacity was scarce. SaaS vendors had better product teams, faster releases, cleaner interfaces and broader benchmarking across customers. The corporate usually did not want to become a software company.

That answer still holds in many places but it does not hold everywhere anymore:

AI changes the economics of internal build.

It compresses some development work.

It accelerates prototyping.

It allows small teams to create internal tools that would previously have required larger engineering teams.

It makes workflow-specific software more plausible.

But this is where companies need discipline:

AI does not remove the cost of ownership.

It does not remove security.

It does not remove maintenance.

It does not remove regression testing.

It does not remove user support.

It does not remove legal accountability.

It does not remove audit.

It does not remove the need to keep the tool alive when the person who built it leaves.

So the right answer is not “build everything”, it is to segment the software estate.

• Buy commodity systems where standardisation is valuable.

• Buy where the vendor has network scale, regulatory depth, security maturity, resilience or superior maintenance.

• Build or customise where the workflow is proprietary, strategically material, data-rich, relationship-rich and poorly served by generic tools.

• Rent frontier intelligence where breadth matters.

• Own the proof where accountability matters.

This is where the corporate’s intangible assets become central. A workflow that depends on the company’s customer trust, supplier history, exception knowledge and risk appetite may be too important to leave entirely inside a vendor interface. Equally, a generic workflow that the vendor performs better and more cheaply should not become a vanity internal build.

The board should not ask whether the company is pro-SaaS or anti-SaaS.

It should ask whether each software decision protects, uses or leaks the operating knowledge that makes the company distinctive.

Falling token prices are not a simple win

The intuitive view is that falling AI prices solve the problem:

They do not.

They change the problem.

Inference prices have been falling quickly, although estimates vary by methodology, model tier and benchmark. A 2025 paper using data from Artificial Analysis and Epoch AI found that the price for a given benchmark-performance level was falling around 5x to 10x per year for frontier models across several benchmark categories. A 2026 paper on inference-service pricing found an approximately 600-fold decline in token prices in its dataset from 2020 to 2026, but also found that flagship models still carried a substantial reasoning premium. This is an academic preprint and should be treated as research evidence, not settled market fact.

Cheaper tokens make more machine work viable and that is the opportunity:

• They allow more experimentation, more automation, more local tools, more digital co-workers and more workflow redesign.

• They may also reduce the cost of errors if the company has good routing, verification and stop conditions.

But cheaper tokens can also expand usage faster than prices fall and that is a risk:

• More workflows become AI-enabled.

• More agents run.

• More internal tools get built.

• More supplier products add AI features.

• More SaaS vendors reprice.

• More employees become heavy users.

• More machine outputs require review.

The result may be lower unit prices and higher total dependency.

This is an enterprise version of Jevons’ paradox: the cheaper the machine work becomes, the more work is likely to be given to machines.

That does not mean the outcome is bad. It means the company must measure the right thing.

Cheaper AI is useful only if it produces more verified value, not merely more machine activity.

There is also a supply-chain question:

McKinsey estimates that data centres may require $6.7 trillion of worldwide capital expenditure by 2030 to keep pace with demand for compute power, including $5.2 trillion for AI-processing loads. McKinsey also sets out a wide scenario range for AI-related data-centre capital expenditure, which is important because the exact demand path is uncertain.

That capital has to earn a return.

If token prices fall sharply and enterprise usage expands, the market may be fine. If token prices fall faster than monetisable demand grows, parts of the AI infrastructure and vendor chain may come under pressure. If vendors subsidise usage to gain adoption, customers may build workflows on economics that later change.

No board should pretend it can forecast the AI infrastructure cycle with precision but it should stress-test its own dependency:

What happens if token prices fall and usage explodes?

What happens if token prices rise for premium reasoning models?

What happens if a vendor changes limits?

What happens if a model is withdrawn?

What happens if the cheap model is not acceptable for regulatory, security or geopolitical reasons?

What happens if a supplier’s AI economics fail?

The risk is not simply that AI becomes expensive, it is that AI becomes cheap enough to embed everywhere before the company has understood what it now depends on.

The AI supply chain changes shape

The first AI supply-chain risk was scarcity:

Could the vendor access GPUs? Could the model provider serve demand? Could the data centre get power? Could the cloud provider allocate capacity? Could the enterprise get enough rate limit to run its workflows?

That risk has not disappeared. McKinsey has separately described the AI data-centre build-out as one of the largest infrastructure build-outs in modern history and points to constraints in power, thermal equipment and long-lead components.

But another risk is now appearing, volatility:

Prices change. Model rankings change. Providers change terms. Open models improve. Closed models cut prices. Some vendors chase enterprise customers. Some may lose margin. Some applications become easier to replace. Some internal tools become more attractive. Some AI-enabled SaaS products become less defensible.

A corporate workflow that depends on one model, one vendor, one orchestration layer, one cloud platform, one data connector or one pricing assumption may look efficient but it may also be fragile.

For each important workflow, it should know:

• Which model is used.

• Which provider serves it.

• Which data are accessed.

• Which tools can be called.

• Which SaaS system sits in the path.

• Which human verifies the output.

• Which fallback exists.

• Which cost assumption makes the business case work.

The AI supply chain is not only the chips and data centres, it is the chain of decisions through which machine work becomes business action.

The CFO version

For the CFO, this is a cost-accounting problem:

The company is building a machine-work cost base. That cost base does not sit neatly in one line item. It appears in software subscriptions, cloud usage, API calls, data platforms, AI-enabled SaaS uplifts, internal engineering, vendor support, cybersecurity, legal review, model evaluation, governance, training, failed pilots and human verification.

The cost is also dynamic.

A seat may be fixed. A token is not. An agent can run more steps. A coding assistant can consume longer context. A legal review can expand across more documents. A supplier negotiation assistant can call more tools. A customer-service workflow can retry or escalate. A reasoning model can use hidden thinking tokens that make actual task cost difficult to predict.

This is why cost-per-token is useful but incomplete, the CFO needs cost per verified outcome:

• Cost per claim triaged correctly.

• Cost per invoice processed without rework.

• Cost per contract reviewed with the relevant exceptions identified.

• Cost per software change accepted after testing.

• Cost per customer query resolved without repeat contact.

• Cost per supplier risk detected before operational failure.

The CFO should also separate four types of savings.

1. Avoided work: a task disappears.

2. Faster work: the task remains but takes less time.

3. Better work: the same process produces fewer errors or better outcomes.

4. Shifted work: the apparent saving in one area reappears as compute cost, review burden, vendor dependency or downstream risk.

Only the first three create value, the fourth creates a better-looking business case.

The CFO’s central question is:

Which AI spend has produced verified value after the full cost of machine work has been counted?

The CEO version

For the CEO, this is an operating-model problem:

The company cannot become AI-capable by adding tools to existing workflows. That may improve individual productivity, but it does not necessarily redesign the company.

The CEO has to decide where the company’s operating knowledge should sit.

• Inside vendor software?

• Inside internal platforms?

• Inside local models?

• Inside data products?

• Inside human experts?

• Inside digital co-workers with controlled authority?

The answer will vary by workflow.

The CEO should start with the parts of the business where the corporate’s intangible assets are strongest: customer trust, supplier knowledge, operational exceptions, pricing judgement, regulatory insight, product judgement and service history.

Those are the workflows where generic software may not be enough.

The CEO should also be careful about theatrical AI. The useful first redesign may not be a glamorous agent. It may be removing three approval loops from procurement, reducing customer exception handling time, routing maintenance cases better, finding claim anomalies earlier, cutting invoice rework or giving frontline teams access to the right operational memory.

The CEO should insist that each digital co-worker has a role, authority and evidence trail.

It should be clear whether the digital co-worker is:

• Assisting.

• Recommending.

• Executing.

• Escalating.

• Approving.

Those are different activities and should not be hidden under one AI label.

The CEO’s central question is:

Which workflows should be redesigned because digital co-workers change the constraint around speed, cost, judgement and accountability?

The Board version

The board does not need to approve every model, gateway, workflow agent or routing rule but it does need to know whether management is turning corporate intangible assets into verified value or merely buying AI activity.

A useful board discussion should cover ten questions.

1. Which workflows have been redesigned, not merely AI-assisted?

2. Which digital co-workers have defined roles, budgets, authority limits and stop conditions?

3. Which AI usage metrics are activity measures rather than value measures?

4. Which AI outputs require human verification, and how much verification work is being created?

5. Which model-routing rules determine when work goes to cheap models, frontier models, local models or humans?

6. Which SaaS tools mainly wrap our own data, workflows, client relationships and supplier knowledge?

7. Which software should we buy, build, retire or replace with workflow-specific tools?

8. Which AI vendors, model providers or orchestration layers create concentration risk?

9. What happens if token prices fall, rise, become less transparent or vary sharply by model type?

10. Which AI benefits have reached the P&L after full cost, risk, rework and old-process removal are counted?

The board should also ask one uncomfortable question:

Are we using AI to redesign work, or to accelerate work that should no longer exist?

That question is rather central because AI can make bad workflows faster:

It can produce more reports that no one needs. It can create more code that no one can maintain. It can generate more analysis than management can absorb. It can preserve redundant controls by making them cheaper to complete. It can automate internal theatre.

The board’s job is not to celebrate usage, it is to demand proof.

The hard truth

The corporate already owns the hard part: it owns the relationships, data, workflows, permissions, operating history, physical assets and judgement that make enterprise AI valuable.

But ownership is not the same as usability:

If those assets are trapped inside fragmented systems, undocumented exceptions, unexamined workflows, vendor interfaces and human memory, AI will not automatically unlock them.

It may simply create another layer of activity.

The next phase of AI will therefore punish two types of company.

1. The first will underuse AI because it treats the technology as a controlled experiment rather than a new form of machine labour.

2. The second will overuse AI because it treats machine activity as transformation.

Both will miss the point and winners will redesign work:

They will route machine labour intelligently.

They will use expensive models where expensive models matter.

They will use cheaper models where cheaper models are enough.

They will build where the workflow is proprietary.

They will buy where the vendor owns real advantage.

They will require decision receipts where accountability matters.

They will count verified value, not tokens.

They will not tokenmax, they will make the workflow measurable.

The weekend takeaway

AI is not asking large corporates to discover whether they own valuable assets, they already do.

The real corporate asset is the operating system of the business: data, workflows, client relationships, supplier relationships, exception history, judgement, permissions and tangible infrastructure.

Digital co-workers make that asset more usable, but only if the work is redesigned.

The danger is not that companies fail to use AI. Many will use it heavily. The danger is that they use AI to accelerate old workflows, burn tokens, deepen vendor dependency and call the activity transformation.

The next software question is no longer simply buy or build, it is:

Which workflow should we redesign?

Which part should a digital co-worker perform?

Which model should do the work?

Which vendor really adds value?

Which decision needs proof?

Which old step can disappear?

The companies that win will not be the ones with the most AI activity, they will be the companies that know which machine work was worth paying for.

Leave a comment

🧠 Techie Corner for the Non-Techies

1. Token

A token is a unit of text or data processed by an AI model. A short prompt uses fewer tokens. A long document, coding agent or multi-step workflow can use many more. Token usage matters because many AI services are priced directly or indirectly by consumption.

2. Tokenmaxing

Tokenmaxing is the habit of maximising or celebrating AI usage as if usage itself proved productivity. It does not. The company may be consuming more machine work without producing more verified value.

3. Verified machine work

Verified machine work is AI output that has been checked, used and connected to a business outcome. A generated draft is not verified machine work. A contract review that correctly identifies the relevant clauses and reduces expert review time may be.

4. Digital co-worker

A digital co-worker is an AI-enabled system that performs part of a workflow. The label is only useful if the company defines its scope, authority, access rights, budget, escalation rules and evidence requirements.

5. Model routing

Model routing sends different tasks to different models or systems. Simple work may go to a smaller or cheaper model. Complex work may go to a frontier model. Sensitive work may stay local. Consequential work may go to a human.

6. Frontier model

A frontier model is a highly capable model near the leading edge of current AI performance. There is no single universal threshold. Frontier models can be valuable, but they are not always the right tool for routine work.

7. Small language model

A small language model is a narrower model that may be cheaper, easier to control and more suitable for bounded workflows. It is not automatically safer or better. It must still be tested against the task.

8. Context window

The context window is the amount of information the model can consider during a task. Larger context can help, but it can also increase cost and noise. More context is not always better.

9. Prompt caching

Prompt caching reduces cost by reusing repeated parts of a prompt or context rather than processing the same material from scratch each time. It is useful where workflows have stable repeated instructions or reference material.

10. Agentic AI

Agentic AI can pursue goals across several steps, often using tools, systems or external data. Its risk depends less on the word “agent” and more on what it can access and do.

11. AI gateway

An AI gateway sits between users or applications and model providers. It can manage routing, logging, cost controls, security, caching and provider selection.

12. Decision receipt

A decision receipt records how an AI-assisted output was produced: what data were used, which model ran, which tools were called, who reviewed the result and whether it was approved. Serious work needs this record.

13. Make versus buy

Make versus buy is the decision whether to build internally or purchase from a vendor. AI lowers some build costs but does not remove maintenance, security, audit, support or accountability.

14. Agentic arbitrage

Gartner uses this term for the risk that AI agents complete tasks across systems and reduce the need for users to interact with traditional SaaS interfaces. The result may pressure parts of enterprise application spending.

15. Jevons’ paradox

Jevons’ paradox is the idea that greater efficiency can increase total consumption by making a resource cheaper or easier to use. Cheaper tokens may increase total AI use.

16. Verification debt

Verification debt is the review burden created when AI generates more output than the organisation can confidently inspect, approve and act on.

17. AI supply chain

The AI supply chain includes model providers, cloud platforms, data centres, chips, power, software vendors, orchestration tools, data connectors and the people who operate them. For corporates, it also includes the workflow dependencies created by AI tools.

18. Workflow redesign

Workflow redesign means changing how work moves through the organisation, not simply adding AI to the existing process. The test is whether steps disappear, decisions improve or outcomes change.

19. Corporate intangible stack

The corporate intangible stack is the combination of data, workflows, customer relationships, supplier relationships, exception history, judgement, trust and permissions that makes the company’s work hard to copy.

20. Cost per verified outcome

Cost per verified outcome measures the full cost of producing a business result, including model use, tokens, cloud, integration, human review, rework, risk and governance. It is more useful than cost per token.

📚 Further reading

• Reuters: Cheaper AI is better: Soaring bills are reshaping how businesses choose models (2026)

  Reuters reports that usage-based AI pricing is making corporate spending unpredictable, that some businesses are turning to smaller and cheaper models, and that routing tools are becoming more important. It also reports examples and estimates from third parties.

• Business Insider: Coinbase’s CEO outlined 5 strategies to keep AI spend low without limiting tokens (2026)

  Business Insider summarises Brian Armstrong’s public comments on Coinbase’s approach to AI spend, including cheaper default models, routing, caching, leaner context and spend visibility.

• Gartner: AI coding costs will surpass average developer salary by 2028 as token consumption surges (2026)

  Gartner argues that token-driven AI coding spend is creating budget strain and recommends model-routing, task classification and disciplined operating models.

• Gartner: $234 billion in enterprise application software spend is at risk from agentic AI (2026)

  Gartner estimates that up to $234 billion of enterprise application software spend could be exposed to agentic arbitrage by 2030, equal to roughly 20% of enterprise application SaaS spending.

• McKinsey: The cost of compute: A $7 trillion race to scale data centers (2025)

  McKinsey estimates that data centres may require $6.7 trillion in worldwide capital expenditure by 2030, including $5.2 trillion for AI-processing loads. The article provides scenario ranges, which matters because demand, efficiency, regulation and supply constraints remain uncertain.

• McKinsey: The $7 trillion data center build-out: How industrials can capture their share (2026)

  McKinsey describes the AI data-centre build-out as a large industrial infrastructure challenge, with constraints around capital, energy and long-lead equipment.

• Mingdeng Du: Tiered Super-Moore’s Law: Price Evolution, Production Frontiers, and Market Competition in Large Language Model Inference Services (2026 preprint)

  This academic preprint studies token pricing across a large dataset and finds large declines in token prices, with important differences by model tier and a continuing premium for flagship reasoning models. It is useful research, but it should be treated as a preprint unless and until independently validated through peer review.

• Hans Gundlach, Jayson Lynch, Matthias Mertens and Neil Thompson: The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference (2025 preprint)

  This paper estimates that the price for a given level of benchmark performance has declined rapidly, around 5x to 10x per year for frontier models across several benchmark categories. It is useful for understanding the direction of inference economics, but benchmark-price declines do not automatically translate into lower enterprise workflow costs.

Last week on Something for the Weekend

Your Cyber Controls Were Built for a Slower Clock

Last week’s article argued that the Five Eyes AI cyber warning is fundamentally about time. AI may not need to invent entirely new attacks. It may only need to make familiar attacks faster than corporate governance can respond.

The central idea was governance latency: the time between recognising a credible risk and executing the decision required to contain it. The article argued that boards should stop asking only whether cyber controls exist and start asking whether those controls can perform quickly enough under pressure.

That same idea carries into this week’s piece: the issue is not whether AI tools exist, the issue is whether the company can redesign, route, verify and govern machine work before the meter becomes the strategy.

The board receives its quarterly cyber report. Multi-factor authentication covers almost every employee. Critical systems are backed up. Endpoint protection is deployed. The incident-response plan was reviewed this year, and the company patches critical vulnerabilities within the service level agreed with the technology team.

The dashboard is green.

Then, the Chief Information Security Officer places a three-page statement in front of the board. It is dated 22nd June 2026 and signed by the heads of the principal cyber security agencies of Australia, Canada, New Zealand, the United Kingdom and the United States.

Not one agency. Six of them.

The Australian Signals Directorate, Canada’s Communications Security Establishment, New Zealand’s Government Communications Security Bureau, the United Kingdom’s National Cyber Security Centre, and the United States’ Cybersecurity and Infrastructure Security Agency and National Security Agency are all saying the same thing.

Artificial intelligence is increasing the speed, scale and sophistication of cyber threats. The time before frontier AI materially changes offensive and defensive cyber capability is not measured in years.

It is measured in months.

The chair looks again at the dashboard. Nothing on it is technically wrong. The controls exist. The policies are current. The audits were completed.

Then, the chair asks a different question:

The room goes quiet.

The backup policy does not say how long it takes to restore the factory. The patching target does not say whether the company can apply an emergency update without stopping production. The identity controls do not say who can revoke every privileged account during a live incident.

The response plan does not say what happens when email, messaging, identity and the board portal are all unavailable at the same time.

The insurance schedule lists policies, limits and deductibles. It does not show how long the company has to notify each insurer, appoint approved specialists, preserve evidence and establish whether one event has crossed several policy classes.

The controls exist.

The problem is whether they can move quickly enough.

That is the uncomfortable implication of the Five Eyes warning.

AI may not need to invent an entirely new form of cyberattack. It may only need to make familiar attacks faster than corporate governance can respond.

The Boardroom point

The Five Eyes statement is not primarily a warning about intelligent machines attacking companies by themselves, it is a warning about time.

AI can help attackers discover weaknesses, study targets, write convincing messages, adapt malicious code, process stolen information and operate across more organisations at once. Defenders can use the same technology to find vulnerabilities, improve software, detect unusual activity and respond more quickly.

Both sides accelerate, but they do not begin from the same position.

An attacker can act when an opportunity appears. A company normally has to assess, discuss, approve, coordinate, document and then act.

That difference creates a new form of exposure:

Governance latency is the time between recognising a credible risk and executing the decision required to contain it. It appears in the gap between a vulnerability being disclosed and the patch being applied. It is the time between an account looking suspicious and its access being revoked, or between a security alert appearing and production being isolated.

It also appears when a supplier is compromised but its connection remains open, when a payment instruction is questioned but the transaction is not stopped, or when a cyber event is detected but insurers are not yet notified.

Those gaps have always existed.

The Five Eyes warning is that AI could make them more expensive by compressing the time on the other side.

The official statement is unusual. It is not a technical advisory about one malicious campaign, one software flaw or one named threat actor. It is signed by the operational leaders of the Five Eyes cyber agencies and addressed directly to boards and senior executives.

It says cyber resilience is integral to continuity, market confidence and long-term value. It asks leaders to understand readiness and accountability, prioritise foundational controls, give cyber leaders sufficient authority and remain actively engaged as the threat changes.

Leaders must be confident that those controls will perform during a real incident.

That changes the test. The old question was:

Do we have the control?

The better question is:

How quickly, reliably and independently can the control perform under pressure?

A backup that has never restored the critical operation is not yet a recovery capability. A cyber-response plan that depends on compromised systems is not yet a plan. A patching policy that takes longer than the exploitation window is a record of good intentions.

A crisis committee without pre-agreed authority is a meeting scheduled inside an attack and a dashboard showing controls as present can therefore be accurate and still misleading. It measures the existence of security.

The Five Eyes warning is about the performance of resilience.

The attack clock has changed

Corporate cyber risk now runs on at least four clocks.

The first is the attack clock. This is the time available between a weakness becoming discoverable and an attacker being able to exploit it. The Five Eyes agencies warn that AI is reducing that interval.

The UK’s National Cyber Security Centre has separately warned organisations to prepare for a vulnerability “patch wave”. Its concern is that increasingly capable AI systems could find weaknesses across large software estates more quickly, exposing years of accumulated technical debt.

The second is the decision clock. This is the time required for the company to decide what to do.

Can the security team isolate a production environment immediately? Can it block a supplier connection, disable a senior executive’s account, stop a payment or take a customer-facing service offline?

Or must it first obtain approval from technology, operations, legal, finance, communications and the executive committee?

That process may be entirely reasonable under ordinary conditions.

During an accelerated attack, it may become the vulnerability.

The third is the recovery clock. This measures how quickly the company can restore a minimum viable operation after systems are lost or deliberately isolated.

The relevant test is not when the backup job last ran. It is when the business can take orders, manufacture safely, dispatch goods, make essential payments, communicate with employees, serve customers, calculate its financial position and meet regulatory obligations.

The fourth is the capital clock. This begins when the technical event starts producing financial consequences.

Cash stops arriving. Emergency expenditure begins. Customers make claims. Contractual penalties accumulate. Suppliers need support.

Insurance notices must be issued. Loss adjusters, forensic specialists and legal advisers must be appointed. Disclosure obligations may arise. Reserves need to be estimated. The company may need liquidity before it knows whether its insurers will respond.

The board should compare all four clocks.

The company does not move at the speed of its fastest security tool.

It moves at the speed of its slowest consequential decision.

Governance latency is now exposure

Boards are used to discussing cyber maturity, they should begin discussing cyber latency.

An organisation can be mature on paper and slow in practice. It can have sophisticated tools but unclear authority. It can detect quickly but decide slowly.

It can isolate systems quickly but recover badly. It can restore technology but not the business process. It can notify insurers but lack the evidence required to allocate the loss.

In each case, the control exists, yet, the delay remains.

Governance latency grows when authority is fragmented:

• The security team may own detection.

• The technology team owns the systems.

• Operations owns production.

• Finance owns payments.

• Legal owns privilege and disclosure.

• Communications owns external statements.

• The board owns oversight.

No one necessarily owns the moment when the company must choose between stopping the operation and allowing the threat to continue. That is not an argument for removing governance.

Speed without control can produce a second crisis.

A badly tested emergency patch can stop a business more effectively than the attacker. An indiscriminate shutdown can damage equipment, interrupt safety systems or destroy evidence. A rushed public statement can create legal and reputational consequences.

The objective is not to act instantly, it is to decide deliberately within the time available.

That requires the difficult decisions to be designed before the incident.

The organisation must know who can isolate the network, stop production, suspend payments, revoke privileged access, move to manual operation, notify customers, invoke the insurance programme and authorise emergency expenditure.

It must also know who can decide that the available evidence is too unreliable to continue.

If those answers are discovered during the incident, governance has already become part of the loss.

The patch wave changes the operating model

Patching sounds administrative until it interrupts a factory, aircraft, hospital, payment platform or customer service.

For ordinary business software, an update may be tested and deployed relatively quickly. For operational technology, medical systems, industrial control equipment and heavily integrated legacy platforms, the process can take much longer.

The company needs to know whether the update is compatible. The vendor may need to certify it. Operations may require a planned outage. Safety teams may need to approve the change.

A customer may need to validate the system.

The company may have no realistic replacement for unsupported technology.

That is why technical debt becomes strategic debt:

A legacy system does not merely cost more to maintain. It can impose a minimum response time on the entire organisation.

AI-assisted vulnerability discovery could expose more weaknesses than security teams and software vendors can remediate through existing processes.

That does not mean every vulnerability will be exploited or that every organisation will face the same pressure. The Five Eyes statement does not provide enough public evidence to quantify the increase in attack frequency or the precise reduction in exploitation time. But the direction of travel is clear enough for management action.

The company needs to identify which systems cannot be patched quickly, which cannot be isolated and which have no supported replacement. It needs to know which patches require extended operational approval and which services depend on suppliers whose patching capability the company cannot observe.

It must also know which critical operations can continue safely while the underlying system is unavailable.

Patching is no longer simply a technology service level, it is part of the company’s operating resilience.

The attacker does not need to be autonomous

The debate about AI-enabled cyber risk is often pushed too quickly towards autonomous machines conducting complete attacks without human involvement.

It is not necessary for the current warning to be serious:

AI can make a human attacker more capable without replacing the human.

It can help analyse an organisation’s exposed systems, translate technical material, produce tailored messages at scale, modify scripts, process stolen documents and identify valuable individuals and suppliers.

It can also suggest alternative attack paths when one fails.

That is enough to change the economics of attack.

The same applies to defence:

AI can help identify unusual behaviour, examine code, prioritise vulnerabilities, summarise incidents and assist overstretched analysts.

But a defensive AI system does not compensate for weak identity controls, unsupported technology, excessive access or untested recovery.

It can help the organisation move faster.

It cannot decide what the organisation is willing to stop, lose or disclose, that remains a management decision.

The control questionnaire is no longer enough

When it comes to the corporate’s insurance, cyber underwriting has often relied on evidence that particular controls are present:

• Does the insured use multi-factor authentication?

• Does it maintain offline backups?

• Does it operate endpoint detection and response?

• Does it patch critical vulnerabilities?

• Does it have an incident-response plan?

Those questions remain useful but they mainly identify nouns when emerging peril lives in verbs.

Not:

Do you have backups?

But:

When did you last restore the critical business process from them, how long did it take and what failed?

Not:

Do you patch critical vulnerabilities?

But:

How quickly can you identify affected assets, test the patch, obtain operational approval and complete deployment?

Not:

Do you have an incident-response plan?

But:

Who has exercised it, which systems were unavailable and what decisions took longest?

Not:

Do you use multi-factor authentication?

But:

How rapidly can you revoke privileged access across cloud, legacy, operational and third-party systems?

Not:

Do you have business continuity arrangements?

But:

Which revenue-producing activities can continue when identity, email, cloud access and central data are unavailable?

The Five Eyes statement creates a challenge for insureds and underwriters alike:

A control can be present at inception and inadequate six months later. The organisation may add new AI systems. The attacker’s capability may improve. A supplier may change its architecture.

A model may be connected to more tools. A business may acquire another company with a weaker estate. A legacy system may become unsupported.

Yet, the risk does not remain static until renewal and that creates a policy-period problem. Insurance is generally priced for a defined period. AI capability and enterprise dependency may change materially within that period.

The annual underwriting snapshot may become less representative of the exposure being carried.

AI is attacking both sides of the balance sheet

The company is likely to use AI to strengthen cyber defence, but it may also introduce AI into customer service, finance, coding, manufacturing, claims, underwriting, procurement and logistics.

Each deployment can create new dependencies and permissions.

An AI assistant that only drafts text presents one exposure. An agent that can search internal systems, call software tools, change records, create code, communicate externally or trigger transactions presents another.

The system could also have access to valuable data previously spread across separate systems and rely on third-party models. It may read untrusted external content, operate continuously and take actions faster than a human reviewer can understand them.

Security therefore has two simultaneous and connected tasks:

1. The first is defending the company against adversaries using AI.

2. The second is defending the company against weaknesses created by its own use of AI.

An attacker does not need to compromise the company’s central network directly. It can merely manipulate the content read by an internal agent, exploit excessive tool permissions, target a model gateway or orchestration layer, or compromise a shared supplier. As a result, it can turn the company’s productivity system into the path of entry.

Trust remains the attack surface

Earlier ‘Something for the Weekend’ articles examined deepfakes, synthetic documents and dashboards that appear reassuring while their underlying evidence becomes less reliable.

The Five Eyes warning extends that problem. AI-enabled attacks are not confined to code, they can exploit authority:

• A convincing message appears to come from a supplier.

• A voice sounds like the CEO.

• A document appears to contain a legitimate update.

• A service account behaves like an approved system.

• An AI agent reads an instruction embedded inside external content.

Instead of a control failure starting with a machine defeating encryption, it can simply start with a person or system accepting the wrong evidence.

That is why identity and access remain central, and advanced security is based on activity-based authorisation if identity is bypassed.

The question is not merely whether someone supplied the correct password or second factor, it is whether the organisation can establish that the person is who they claim to be, the device is trustworthy, the request is consistent with authorised behaviour, the system has not been manipulated and the action remains within delegated authority.

The evidence must also be capable of reconstruction later: trust that cannot be verified becomes a source of operational leverage for the attacker.

Security by design becomes a commercial issue

The Five Eyes agencies call for secure-by-design and secure-by-default practices. Those phrases can sound like product-development doctrine but they are also commercial allocation mechanisms.

If products are insecure by default, customers carry more of the cost of making them safe. If updates are difficult to apply, customers carry more of the operational risk.

If logs are incomplete, customers and insurers carry more uncertainty after an incident. If systems are connected unnecessarily, the insured carries a larger attack surface.

If a vendor cannot explain its dependencies, the whole supply chain carries hidden accumulation:

• The customer may pay for the product.

• The insurer may pay part of the loss.

• The reinsurer may absorb the accumulation.

• Capital-market investors may ultimately carry the tail.

Design decisions made by a software vendor can therefore affect capital far beyond the vendor’s own balance sheet.

The Five Eyes statement is addressed not only to user organisations but also to vendors.

Cyber resilience cannot be created entirely by asking every customer to compensate for insecure products through more controls, more monitoring and more insurance.

Some risk has to be removed at source.

The CFO version

For the CFO, this is not principally a cyber-budget question, it is a question about the speed at which a technical event becomes a financial event. The CFO should map the path from compromise to cash:

• Which systems generate invoices, release payments, calculate payroll and support order fulfilment?

• Which systems provide the evidence required for insurance claims?

• Which suppliers become financially distressed if payments stop?

• How much liquidity is required if operations are interrupted for three days, two weeks or two months?

• When does the company breach a covenant?

• When do customers gain termination rights? When do contractual penalties begin?

• Which losses sit below deductibles?

• Which emergency costs require prior insurer consent?

The company may have enough insurance and still run out of liquidity before the claim is resolved. It may restore the systems and still lose revenue because customers no longer trust them. It may contain the attack and still face litigation, regulatory action, forensic costs and supplier support.

The CFO should also challenge the economics of technical debt. A legacy system may appear cheaper because replacement has been delayed. But its true cost includes slower patching, more manual work, limited logging, expensive specialist support, longer outages, greater dependence on individual employees and higher uncertainty for insurers.

A system that cannot be secured or recovered within the company’s risk tolerance is not merely old, it is an unfunded contingent liability.

The CFO’s central question is:

How much financial exposure accumulates during every hour in which the company detects the problem but cannot yet act?

The CEO version

For the CEO, the Five Eyes warning is about organisational authority. The technology team cannot solve a crisis that requires the business to choose what it is prepared to stop. A cyber event can force decisions across production, customers, employees, regulators, investors and suppliers.

Those decisions belong to management.

The CEO should establish the minimum operation the company must preserve. That may include safe production, emergency customer service, cash control, essential supplier payments, employee communication, regulatory reporting and evidence preservation.

The company should then rehearse how to maintain those functions without its normal digital environment. This is not a return to paper and typewriters for everything. The goal is to define a deliberate degraded mode.

The CEO also needs to remove ambiguity from crisis authority:

The person who can detect an attack may not have authority to interrupt the business.

The person with authority may not understand the technical urgency.

The gap between them is governance latency. The CEO must decide in advance which actions security leaders may take immediately, which require executive approval and which must never be automated.

The company should know where friction is protective and where it is merely delay.

The CEO’s central question is:

Can this organisation make a difficult operational decision at the speed of the threat without losing accountability for the decision?

The Board version

The board does not need to know how to write malicious code or configure an identity platform, but it does need to know whether management’s cyber controls can produce a business decision before the available time expires.

A useful board discussion should cover ten questions:

1. Which critical vulnerabilities, systems or dependencies impose the longest response time on the company?

2. How quickly can management move from a credible warning to an authorised operational decision?

3. Who can isolate systems, revoke privileged access, stop payments or suspend production without waiting for the full executive committee?

4. Which critical services depend on unsupported, difficult-to-patch or poorly understood technology?

5. When did the company last restore a minimum viable operation from backups under realistic conditions?

6. Can the company operate when identity, email, cloud access, collaboration tools and the board portal are unavailable simultaneously?

7. Which third-party platforms, model providers, security vendors and software components create common points of failure?

8. How would one severe event move across cyber, crime, property, liability, business interruption and directors’ and officers’ coverage?

9. What evidence would the company need to establish the cause, scope and financial value of the loss?

10. Which assumptions in the company’s risk appetite, insurance programme and recovery planning would fail if the threat changed materially within months?

The board should receive evidence across four dimensions.

• The first is control performance. The question is not whether the control exists, but when it last worked under test, how long it took and what failed.

• The second is decision performance. The issue is not whether escalation routes are documented, but whether the accountable person can act with sufficient authority and information.

• The third is recovery performance. The relevant evidence is not whether data were backed up, but whether the critical business process can be restored within tolerance.

• The fourth is capital performance. The board needs to understand not merely whether insurance is purchased, but how liquidity, coverage, claims evidence and loss aggregation work during the period before recovery.

The board should also be cautious about artificial precision. The Five Eyes statement does not provide a quantified probability of attack, a named frontier model, a defined capability threshold or a forecast of insured loss. Management should not convert “months, not years” into a fabricated percentage or a false countdown.

The statement is a strategic warning and the correct response is not panic, it is to test whether the organisation’s assumptions remain valid under a faster clock.

The board’s hardest question is:

Are our controls designed to satisfy the governance process, or to preserve the business when the governance process itself is under pressure?

The hard truth

Most companies do not discover that a control is weak when the policy is written, they tend to discover it when the control is needed.

For instance, the backup exists but restores too slowly. The patch is available but cannot be deployed. The emergency account exists but no one remembers where the credentials are held. Or, the manual process exists but the employees who knew it have left. The insurance policy exists but notice was late. The incident plan exists but everyone needs the compromised identity system to open it.

Yet, the dashboard remains green because it measures declared controls rather than demonstrated resilience.

AI does not create every one of these weaknesses, but it exposes them much faster: that is the significance of the Five Eyes warning.

The agencies are not saying that every company will face a devastating autonomous attack within months, but they are saying that the assumptions beneath cyber resilience may become outdated within months.

And, that is a different warning, a more difficult one. Whilst a named vulnerability can be patched and a named threat actor can be tracked, a changing capability environment requires the organisation to keep testing whether its controls, authority and capital still fit the peril.

The weekend takeaway

The cyber industry has spent years helping companies see more: more alerts, telemetry, threat intelligence, dashboards and risk scores.

The Five Eyes statement asks a harder question:

Can the company act on what it sees before the opportunity to act disappears?

That is not just a technology question alone, it is a question of authority, operations, liquidity, insurance and trust. Attackers can use AI to compress reconnaissance, targeting, exploitation and adaptation. Defenders can use AI to detect and respond more quickly.

But AI does not remove the board’s responsibility to decide what must stop, what must continue and who has authority when the evidence is incomplete. In fact, the company with the most cyber controls may look secure but the company whose controls have been tested, timed and connected to operational decisions is more likely to be more resilient.

For boards, controls were built for a particular threat clock, the Five Eyes agencies are warning that the clock is changing.

The time to discover whether your resilience is real is before the attacker does.

Leave a comment

🧠 Techie Corner for the Non-Techies

1. Five Eyes

Five Eyes is an intelligence and security partnership comprising Australia, Canada, New Zealand, the United Kingdom and the United States. The June 2026 statement was signed by the heads of the principal cyber security bodies from those countries, including both the NSA and CISA in the United States. The statement is public evidence of their shared assessment. It does not disclose the classified or private information that may sit behind that assessment.

2. Frontier AI

Frontier AI generally refers to highly capable models near the leading edge of current development. There is no single universally accepted technical threshold that makes a model “frontier”.

3. Governance latency

Governance latency is used in this article to describe the time between identifying a credible risk and executing the organisational decision required to address it. It is not a formal cyber standard or accounting term.

The delay may arise from unclear authority, fragmented ownership, legal review, operational constraints or weak evidence.

4. Exploitation window

The exploitation window is the period between a vulnerability becoming known or discoverable and its successful use by an attacker. The period varies by vulnerability, attacker, target and disclosure process.

AI may shorten parts of this process by assisting research, coding, reconnaissance or target selection.

5. Vulnerability patch wave

The UK National Cyber Security Centre uses this phrase for a possible increase in software patches as AI systems become better at finding vulnerabilities. It does not mean that every discovered weakness will be severe or exploited.

The operational challenge is that organisations may need to evaluate, test and deploy more security updates within shorter periods.

6. Technical debt

Technical debt is the future cost and constraint created when organisations retain old, unsupported, poorly documented or difficult-to-change systems.

The phrase is broader than obsolete hardware. It can include fragile integrations, unsupported software, inadequate logging and processes that depend on a small number of people.

7. Attack surface

The attack surface is the collection of systems, identities, interfaces, connections and processes through which an attacker may try to gain access or cause harm.

Reducing the attack surface means removing unnecessary exposure and access rather than trying to defend every possible connection indefinitely.

8. Defence in depth

Defence in depth uses several independent layers of protection: if one control fails, another may prevent or limit the loss. Examples include identity controls, network segmentation, endpoint protection, monitoring, backups and manual recovery procedures.

The layers should not all depend on the same provider or system.

9. Secure by design

Secure by design means considering security during the design and development of a product rather than adding it later.

This can include reducing default privileges, minimising exposed services, producing reliable logs and making security updates easier to apply.

10. Secure by default

Secure by default means that a product begins with protective settings already enabled.

The customer should not need specialist knowledge to avoid an insecure configuration.

11. Zero-day vulnerability

A zero-day is a vulnerability for which the affected organisation or vendor has no effective patch available when exploitation begins or becomes known. The term is sometimes used imprecisely.

Not every newly discovered vulnerability is a zero-day, and not every zero-day is actively exploited.

12. Operational technology

Operational technology monitors or controls physical processes and equipment. Examples include manufacturing controls, building systems, utilities and industrial machinery.

Patching operational technology can be slower and more complex because changes may affect safety, certification and physical production.

13. Identity and access management

Identity and access management controls who or what can enter systems and what actions they may take. It applies to employees, contractors, software services, machines and AI agents.

Strong identity control includes authentication, least privilege, regular access review and rapid revocation.

14. Least privilege

Least privilege means giving a person or system only the access required for its task. An AI agent that can read documents does not necessarily need authority to change records, send payments or modify production systems.

Reducing authority reduces the potential blast radius of compromise or error.

15. Agentic AI

Agentic AI systems can pursue goals through several steps, including planning, reading information, selecting tools and taking actions.

Their risk depends less on the label “agent” than on the authority, data and systems connected to them.

16. Prompt injection

Prompt injection occurs when untrusted content influences an AI system’s instructions or behaviour. It is sometimes compared with social engineering for machines.

The risk becomes more consequential when the AI system can use tools or act in external systems.

17. Model Context Protocol

Model Context Protocol, usually shortened to MCP, is a protocol used to connect AI systems with tools, data and services. It can simplify integration, but it can also create a security boundary requiring strong authentication, permissions, validation and monitoring.

18. Black-start capability

Black-start is borrowed from power-system terminology. In this article, it refers to the ability to restart a minimum viable business operation without relying on the systems that normally support it.

It may involve isolated accounts, alternative communications and manually controlled processes.

19. Cyber accumulation

Cyber accumulation occurs when one event creates losses across several insureds, policies or lines of business. Common technology providers, software components and cloud platforms can connect organisations that otherwise appear unrelated.

20. Event definition

Event definition determines which losses are treated as arising from one occurrence or several occurrences. This affects insurance limits, deductibles, reinsurance recovery and ILS triggers.

Cyber events can be difficult to define because attacks may unfold across systems, victims, actors and time.

21. Minimum viable operation

A minimum viable operation is the smallest set of business activities that must continue or be restored to preserve safety, liquidity, legal compliance and essential customer service.

It is not necessarily the same as restoring every system.

📚 Further reading

• Five Eyes cyber security agencies: The AI Shift in Cyber Risk: Why Leaders Must Act Now (2026)

  The statement says AI is accelerating the speed, scale and sophistication of cyber threats and that frontier capabilities could transform offensive and defensive cyber activity within months rather than years. It calls for foundational security, defence in depth, faster patching, stronger identity controls, incident preparation and deliberate use of AI for defence. The statement does not disclose the underlying intelligence, identify a specific model or quantify the expected increase in loss.

• UK National Cyber Security Centre: Retaining Defensive Advantage in the Age of Frontier AI Cyber Capabilities (2026)

  An explanation of why AI-assisted vulnerability discovery may expose organisations with weak security foundations. It argues that defenders need to raise baseline security and use AI carefully rather than assuming AI tools will compensate for technical debt.

• UK National Cyber Security Centre: Preparing for a Vulnerability Patch Wave (2026)

  Guidance on the organisational consequences of faster vulnerability discovery. It focuses on asset knowledge, patching capacity, legacy technology and the operational ability to process a higher volume of security updates.

• UK National Cyber Security Centre: Preparing for Severe Cyber Threat: Why Leaders Must Act Now (2026)

  Board-level guidance on preparing for cyber events capable of producing substantial operational disruption. It distinguishes ordinary compliance from readiness for severe incidents and emphasises rehearsed recovery and executive authority.

• UK National Cyber Security Centre: Impact of AI on Cyber Threat from Now to 2027 (2025)

  The NCSC’s assessment of how AI may affect reconnaissance, social engineering, vulnerability research, exploit development, malware and stolen-data processing. It expects AI primarily to make existing attack techniques faster and more effective in the near term. The report does not conclude that fully autonomous advanced attacks are already the norm.

• CISA, NSA, ASD’s Australian and International Partners: Careful Adoption of Agentic AI Services (2026)

  Joint guidance on the risks created when AI systems can plan, use tools and take actions with limited supervision. It recommends incremental adoption, layered controls, explicit accountability, continuous monitoring and human oversight.

• World Economic Forum: Global Cybersecurity Outlook 2026

  Survey-centric analysis highlighting that 87% of security executives view AI-related vulnerabilities as their fastest-growing threat vector. The report charts the macroeconomic impacts of data leaks, geopolitical fractures, and widespread cyber fraud.

• National Institute of Standards and Technology: Cybersecurity Framework 2.0 (2024)

  A general framework structured around governing, identifying, protecting, detecting, responding and recovering from cyber risk. It is voluntary guidance and does not establish that an organisation using it will withstand a severe incident.

• National Institute of Standards and Technology: Artificial Intelligence Risk Management Framework and Generative AI Profile (2023/2024)

  Frameworks for governing and managing risks arising from AI design, deployment and use. They cover governance, measurement, third-party risk, testing, monitoring and accountability.

• Cybersecurity and Infrastructure Security Agency: Secure by Design

  Guidance encouraging technology manufacturers to make security a core product requirement and to provide protective defaults. It is relevant to the Five Eyes statement’s position that security costs and responsibilities should not fall solely on customers.

• Swiss Re Institute: Research on AI risk and cyber accumulation (2026)

  Swiss Re has examined how AI can create interconnected risks across cyber, fraud, liability and business interruption, as well as how dependence on major cloud providers can create accumulation. These publications are insurer research rather than neutral market-wide loss data.

• Dr Henri Winand: When No One Believes the Dashboard: The Trust Recession in a Deepfake World (2025)

  Assesses corporate internal trust failures, arguing that executive metrics become empty targets when deepfakes and AI synthetic evidence decouple a status report from its actual ground truth.

• Dr Henri Winand: Single Points of Failure You Don’t See (2025)

  An examination of hidden dependencies in identity, payments, logistics and operational systems. Its central recommendation was to know the critical dependencies, limit their blast radius and rehearse recovery without them.

• Dr Henri Winand: When Deepfakes Hit Treasury (and Insurance Notices Too)(2025)

  An article on synthetic trust, payment controls and the ambiguity between cyber, crime and social-engineering coverage.

• Dr Henri Winand: When Synthetic Trust Starts Leaking Internally (2026)

  An examination of how AI-generated documents can degrade financial signals without creating one easily identifiable major incident.

• Dr Henri Winand: Securing the Machine That Decides (2026)

  An article on agentic AI, delegated authority, prompt injection, tool access, logging, human control and kill-switch design.

• Dr Henri Winand: AI as an Aggression Multiplier (2026)

  When machines optimise both attack and defence, cyber risk stops being episodic and becomes a volatile, economically calibrated campaign.

• Dr Henri Winand: The Technology and AI Competitive Trap (2026)

  An examination of why voluntary safety commitments can weaken when technology developers, states and companies face strong incentives to accelerate.

Last week on Something for the Weekend

The Great Ownership Transfer could become a major AI productivity platform. As millions of established businesses change hands, new owners could preserve their customers, employees and specialist knowledge while applying AI, robotics, better software and capital to improve them.

The real advantage would not be access to a model, but a repeatable operating system for capturing knowledge, redesigning workflows and transferring learning across acquisitions. The outcome will depend on governance: whether the platform builds stronger companies and shares the gains, or simply extracts knowledge, cuts jobs and prepares businesses for resale.

The owner is 68. She started the company thirty-seven years ago with two technicians, a second-hand van and a loan secured against the family home.

Today, the business employs 83 people. It repairs industrial pumps, fabricates replacement components and maintains machinery for food-processing plants across three states. The company is profitable. Its customers renew. Its employees know the machines, the plants and the people who run them. Its reputation is stronger than its website.

The accounts are prepared once a year. Some jobs are still scheduled on a whiteboard. The sales pipeline lives partly in a spreadsheet and partly inside the owner’s head. The youngest engineer has built a small AI tool to search old maintenance reports, but the company has no formal data architecture, no AI policy and no one responsible for technology.

The owner’s children have careers elsewhere. The general manager could run the company but cannot afford to buy it. A private-equity firm is interested, but the business is below its usual transaction size. A strategic buyer could acquire it, remove duplicated costs and consolidate the operation into a regional platform.

The owner worries that the name will disappear. The employees worry that their jobs will disappear. The customers worry that the people who understand their equipment will disappear.

So, the owner waits.

That is not one succession problem. It is one example of a much larger transfer of productive assets, customer relationships, operating knowledge and local employment that is beginning to move from one generation to the next.

McKinsey calls it the “Great Ownership Transfer”: USD5 trillion in enterprise value. Its analysis estimates that millions of small and medium-sized US businesses could face an ownership transition as their proprietors retire. Many are too small for conventional private equity. Many lack an obvious family successor. Many contain valuable operating knowledge but weak digital infrastructure.

Many will be sold. Some will be passed to employees or family members. Some will close.

Then, a different buyer arrives and it does not see an old pump-maintenance company.

• It sees a customer network, a skilled workforce, thirty-seven years of failure data, thousands of maintenance records, a recurring service book and a physical distribution footprint.

• It sees a platform on which artificial intelligence, robotics, better software, procurement scale and modern capital can be applied.

The buyer does not begin by asking how many employees can be removed. It asks how much more valuable the business could become if its people were given better machines.

Then, it asks the more unsettling question:

The Boardroom point

The largest AI opportunity may not be another chatbot. It could just be the transfer and rebuilding of the existing economy.

The Great Ownership Transfer is usually described as a succession problem. Older owners retire. Businesses need buyers. Communities need continuity. Employees need protection. Capital needs a way to complete smaller transactions economically.

However, the same transition may also create an unusual industrial opportunity. A new owner can acquire a functioning business at the moment when its operating model is already due for renewal.

The company has loyal customers, experienced employees, positive cash flow, specialist knowledge, local market power and trusted supplier relationships. It also has underdeveloped data, manual processes, fragmented software, weak succession planning and limited capital for modernisation.

That mix is important.

Often, the difficult part of building a business is not creating the software. It is winning customers, developing trust, learning the exceptions, hiring people who understand the work and surviving long enough to accumulate useful knowledge.

Those assets already exist inside many businesses approaching succession. The opportunity is therefore not simply to start new AI-native companies and wait for them to replace the old economy.

It may be to acquire parts of the old economy and make them AI-capable.

Jeff Bezos is probably assembling the pieces of precisely such a model.

The Wall Street Journal reported in March 2026 that Bezos was in early talks to raise up to $100 billion for what investor materials reportedly described as a manufacturing transformation vehicle. According to that reporting, the proposed fund would acquire manufacturing businesses and use artificial intelligence to accelerate automation and improve their economics. The “Manufacturing Transformation Vehicle” is likely to target larger businesses. This may offer an entry point for mid market PE firms rolling up assets for larger buyers.

Separately, Bezos and Vik Bajaj are co-leading Prometheus, an AI company focused on engineering and the physical economy. Prometheus has said it wants to build an artificial general engineer capable of helping humans design, simulate and manufacture complex physical products. The company announced in June that it had raised $12 billion at a valuation of approximately $41 billion.

The combination is notable:

• Capital to acquire companies.

• AI designed for engineering and manufacturing.

• A founder with experience operating physical systems at enormous scale.

• A prospective portfolio in which technology can be deployed repeatedly.

Amazon provides the operating precedent. By June 2025, Amazon said it had deployed its one-millionth robot across a network of more than 300 facilities. It has developed AI systems for demand forecasting, routing, warehouse movement, inventory handling and increasingly natural-language interaction with robotic systems.

Its newer systems include machines that move heavy loads, manipulate objects and assist people working inside fulfilment centres. Amazon also says more than 700,000 employees have participated in its education and skills programmes globally.

So, Amazon has spent decades learning how to combine software, data, robotics, physical infrastructure, process design and human labour across a large distributed operation. Very few organisations have comparable experience.

A Bezos-backed acquisition platform or other similarly technologically enabled mid market approaches could therefore bring more than capital and a licence for an AI model. It could bring a playbook for integrating intelligence with the physical economy.

If that worked, the platform would not merely own a collection of factories. It would own a repeated method for making established businesses more productive.

That method could become more valuable with every acquisition.

Last week, waste became a management problem

Last week’s Something for the Weekend argued that the greenest AI is the one that wastes least.

The central measure was not tokens per pound or output per kilowatt. It was:

Verified business value per unit of total resource consumed.

That argument becomes more important when AI moves from an employee’s browser into the operating system of a company.

But a business acquisition creates a harder test than a software pilot. The buyer cannot claim success because employees generated more documents or used more prompts.

It has to demonstrate that the company produced more, sold more, made fewer errors, reduced downtime, improved quality, shortened working-capital cycles, served customers faster, created safer jobs, raised wages or developed products more quickly.

AI adoption becomes accountable when it ‘owns’ the profit and loss account.

The Great Ownership Transfer may therefore create an unusual laboratory for measuring whether AI genuinely improves the real economy.

Not whether a model passed a benchmark. Not whether a board attended an AI workshop. Not whether employees received licences.

Whether a transferred business became more productive, resilient and valuable without destroying the knowledge that made it worth buying.

The business is not obsolete

A common mistake in technology is to confuse an old interface with an obsolete business.

A company can use outdated software and still possess valuable capabilities. The pump-maintenance company could have poor data systems but exceptional knowledge of why a particular machine fails after a certain cleaning cycle. A small aerospace supplier can use manual inspection records but understand tolerances that are not properly captured in its customer’s engineering model.

A regional logistics company could have weak route-planning software but know which deliveries fail when a particular road floods. A specialist insurer could run an old policy administration system but possess decades of claims judgement in a narrow class of risk.

The opportunity is not to discard that knowledge because it is badly formatted. It is to capture, organise and combine it with better tools.

That distinction separates transformation from liquidation.

The business being transferred should already contain the difficult assets: customer trust, tacit knowledge, skilled labour, supplier relationships, regulatory approvals, physical sites, a known product and a record of what works.

The AI system is not the business. It is a way of helping the business learn, decide and operate more effectively.

The missing middle is also a data problem

McKinsey’s report highlights an awkward mismatch: a large proportion of businesses approaching transition have enterprise values below the level at which many institutional investors can deploy capital efficiently. Each transaction requires finding the seller, assessing the accounts, evaluating management, reviewing contracts, understanding customers, conducting legal and tax work, arranging finance, negotiating terms, transferring ownership and supporting the new operator.

For a small transaction, much of this work costs nearly as much as it would for a larger one.

The result is the missing middle: businesses too substantial for an informal handover, too small for conventional private equity, too complex for an inexperienced buyer and too important to their employees and communities simply to close.

AI could reduce some of that transaction friction. It may help structure unorganised financial information, search contracts, identify missing documentation, compare operating metrics, prepare initial diligence questions, map customer concentration, review maintenance and quality records, build post-acquisition operating plans and match sellers with potential buyers.

But that does not make diligence automatic:

• A model cannot confirm that a customer will remain after the founder leaves.

• It cannot inspect a machine merely by reading the depreciation schedule.

• It cannot establish the quality of management from a polished information memorandum.

• It cannot replace legal, financial, environmental, operational or technical judgement.

However, it may reduce the cost of organising the evidence on which that judgement depends which is important when transaction economics are constrained by scale.

Therefore, the first productivity gain could occur before the business is acquired: AI could make the ownership transfer itself easier to execute.

Acquisition is not the same as transformation

Buying old companies and adding AI to the investment memorandum is ‘easy’. Changing how they operate is not.

Private equity has spent years describing portfolio-company digitisation, data transformation and technology-enabled value creation. The results have varied.

A company cannot be made AI-native by buying licences, hiring a consultant, creating a dashboard, uploading documents into a chatbot, adding an AI paragraph to the hundred-day plan or renaming cost reduction as automation.

The Markets Group and Brownloop paper describes an AI illusion in private equity. Its central claim is that using the same general-purpose models as every competitor does not create a durable advantage.

Generic models may well improve productivity. However, if every firm can access similar models and public information, the edge must come from elsewhere:

• Proprietary data.

• Institutional memory.

• Workflow integration.

• Governance.

• Decision records.

• Learning accumulated across repeated use.

The same logic applies to an acquisition platform.

The durable asset would not be access to Prometheus or any other model since competitors could eventually gain access to models with similar capabilities.

The durable asset would be the operating system built around the model: how businesses are selected, how their knowledge is captured, how their processes are mapped, how employees are involved, how data is cleaned, how permissions are assigned, how workflows are redesigned, how quality is tested, how savings are measured and how lessons from one acquisition improve the next.

That is where a collection of companies could become a platform.

The portfolio can learn

Picture ten companies making different industrial components. Each has its own maintenance records, supplier history, quality-control system, engineering files and production constraints.

A conventional holding company could manage them as separate investments. A stronger platform would build a common intelligence layer while preserving the differences that matter.

• One company learns that a certain vibration pattern predicts a bearing failure.

• Another identifies a supplier batch associated with premature cracking.

• A third develops a better sequence for machining a difficult alloy.

• A fourth improves energy consumption by changing when equipment is operated.

• A fifth discovers that a particular inspection step catches most defects but two others add little value.

The lessons cannot always be transferred directly. Different machines, materials and customers behave differently. A correlation in one plant may be meaningless in another. Confidentiality, competition law, customer restrictions and intellectual-property rights may limit what can be shared.

Yet a platform can still develop better methods for structuring maintenance records, testing predictions, measuring downtime, routing work, deciding when a human must intervene and integrating new acquisitions.

The system becomes more capable not merely because it owns more data. It becomes more capable because it has repeated the work of turning messy industrial reality into usable organisational knowledge.

That learning curve is difficult for a new competitor to reproduce.

The roll-up changes

Traditional roll-ups often create value through familiar mechanisms such as centralised purchasing, shared finance, common technology, consolidated management, cross-selling, lower overhead, stronger negotiating power, debt and multiple arbitrage.

AI adds another possible mechanism:

A platform could buy businesses valued as isolated local operators and increase their value by connecting them to a shared intelligence system.

The first acquisition helps build the system. The second benefits from what the first taught. The tenth enters a much more mature integration process. The hundredth may receive tools, benchmarks and operating methods that no standalone buyer could justify developing for one small company.

That is the attractive version.

There is also a less attractive one. The platform could acquire fragmented businesses, extract their data, standardise their operations, remove local management, reduce labour, increase leverage, concentrate market power and sell the enlarged group at a higher multiple.

The technology does not determine which version occurs, but ownership, incentives and governance do.

Who owns the intelligence?

The ownership question becomes more important once the platform begins to learn from acquired businesses.

Suppose a company’s employees have accumulated twenty years of specialist knowledge. That knowledge appears in documents, emails, repair notes, drawings, customer conversations, workarounds, judgement, habits and stories told between experienced and younger workers.

The buyer converts part of that knowledge into a structured AI system.

Who owns the resulting asset?

Legally, the answer will usually depend on employment terms, intellectual-property provisions, customer contracts, database rights, privacy law and the architecture through which the information was collected.

But economically, the question is broader.

The seller receives a purchase price. The investor owns the platform. The model provider receives fees. Management may receive equity.

What do the employees receive for helping convert their accumulated experience into a scalable machine asset?

They could receive more secure employment, better tools, higher wages, promotion, training, profit sharing or equity.

Or, they could receive redundancy.

A platform that treats employees merely as temporary containers of extractable knowledge will create resistance and destroy trust. Employees will be rationally reluctant to document the expertise that could be used to remove them.

A platform that allows people to participate in the productivity they help create has a better chance of obtaining the knowledge required to create it.

This is not only merely a fairness argument, it is an operating argument.

The machine cannot learn what the workforce refuses to reveal.

The employment bargain

Jeff Bezos argued at VivaTech that AI could create labour shortages rather than make humans redundant.

AI may increase demand by making invention and production cheaper. It may create new companies, products and occupations. It may also reduce labour requirements in existing tasks and companies.

Both can occur at the same time.

The relevant question for an ownership-transfer platform is therefore not whether AI will create or destroy jobs, it is:

A credible bargain might contain several principles.

1. Do not begin with an arbitrary headcount target

If the business case assumes that a fixed percentage of employees must leave before the work has been redesigned, management is not discovering productivity. It is allocating a predetermined cost reduction.

Of course, some roles may disappear, others expand, and new ones may emerge. The decision should follow evidence about the redesigned operation, not precede it.

2. Share part of the gain

When workers help train systems, redesign processes and improve productivity, the resulting value should not flow exclusively to the acquisition vehicle.

Profit sharing, performance participation, employee equity or funded training can connect the workforce to the upside.

The correct structure will vary by company and jurisdiction. The principle is constant: people should have a route to benefit from making their employer more productive.

3. Fund mobility before displacement

Training offered after a role disappears is often too late.

A serious platform would identify adjacent roles while the transformation is being designed. An experienced machine operator could become a robotics technician, quality-system supervisor, process trainer, data steward, customer-solutions specialist, maintenance planner or human reviewer for automated decisions.

Not everyone will make that transition. Training does not guarantee a new role.

However, the platform should be able to show that it invested in mobility before relying on redundancy.

3. Preserve local operating authority where it creates value

Centralisation can improve finance, technology, procurement and compliance. It can also destroy local knowledge.

The person closest to the machine, customer or production line often sees problems that a remote system misses. Therefore, the objective should not be maximum central control. It should be the correct distribution of authority between the employee, local manager, AI system, central platform and board.

4. Make the system contestable

Employees need a route to challenge an AI-supported instruction or decision.

The machine may optimise the wrong metric or it simply misunderstands an unusual customer requirement. It could also recommend a maintenance delay that looks efficient but creates safety risk.

A human should not be penalised for escalating a material concern merely because the system was confident.

5. Measure job quality, not only job count

A business could retain the same number of people while making their work worse.

Technology can increase surveillance, work intensity and algorithmic pressure. It can also remove repetitive physical tasks, improve safety and expand responsibility.

A serious employment scorecard should examine pay, hours, predictability, safety, training, autonomy, promotion, work intensity, employee turnover and employee participation in value creation.

The number of jobs alone is not enough.

Amazon is the proof and the warning

Amazon demonstrates that AI and robotics can operate at physical scale. It also demonstrates why the employment design cannot be treated as an afterthought.

The company has deployed robotics across a vast fulfilment network. Its systems can move inventory, optimise routes, forecast demand and increasingly respond to natural-language instructions.

Amazon says these technologies remove repetitive and physically demanding work while creating new technical roles. It has also announced extensive training and apprenticeship programmes.

Those are all material capabilities.

At the same time, Amazon has faced criticism over warehouse injury rates, work intensity, algorithmic management, labour relations and employee surveillance.

Those issues are not incidental to a thesis about transforming work. They are central to it.

A productivity platform can create more output per worker while transferring risk, pressure and control onto the worker. It can reduce heavy lifting while increasing the pace at which remaining tasks must be completed. It can create skilled robotics jobs while narrowing autonomy elsewhere. It can offer training while still displacing people who cannot access or complete it.

The lesson is not that physical AI is inherently good or bad, it is that operating design determines how the productivity gain is distributed.

Amazon provides evidence that large-scale physical AI is possible. It does not settle the question of what the ideal human settlement around it should be.

Walmart’s builders show the other side

The transformation does not have to be imposed only by engineers working from headquarters.

Walmart says its internal Code Puppy platform enables non-technical employees to build tools that previously required technical support.

Its Global Chief Technology Officer, Suresh Kumar, said Walmart had as many non-technical associates able to do technology-enabled work as it had engineers.

That statement is Walmart’s own description of its workforce capability, not an independently audited measure. Nevertheless, it illustrates an important shift.

AI can move the ability to design software-like workflows closer to the person who understands the work: the warehouse supervisor, buyer, store manager, finance analyst, maintenance engineer or customer-service employee.

That could be particularly valuable inside transferred businesses.

Small companies often lack large IT teams. Their employees nevertheless understand where work is duplicated, where information is lost and which exceptions consume time.

Giving them controlled tools to redesign those processes could produce better results than importing a central team that does not understand the business.

Walmart also illustrates the governance problem. When many employees become builders, traditional security and software-development controls are no longer be sufficient.

The company says it addresses this through approved models, shared internal platforms, permission controls and embedded governance.

The principle is relevant to any acquisition platform:

Every employee should not receive unrestricted authority over every model, dataset and operational system.

The platform needs to make safe building easier than unsafe building.

The service economy is not excluded

The reported Bezos vehicle has been associated primarily with manufacturing which is understandable. Prometheus is focused on engineering and the physical world. Bezos and his team bring substantial operational experience in logistics, robotics and infrastructure.

However, the broader ownership-transfer thesis extends far beyond factories.

It applies to accounting firms, insurance agencies, engineering consultancies, maintenance businesses, property managers, dental groups, veterinary practices, legal services, freight brokers, recruitment firms, testing laboratories, specialist distributors, construction services and home-care providers.

These businesses will contain fewer robots but similar ingredients: recurring customers, skilled judgement, fragmented records, manual administration, weak succession planning and labour constraints.

AI could improve scheduling, document preparation, customer support, knowledge retrieval, quality control, pricing, compliance, sales administration, claims handling, inventory, work allocation, forecasting and professional review.

However, service businesses create a different risk. Their product may be judgement itself.

Automating a factory movement and automating legal, medical, financial or insurance advice are not equivalent. The cost of an error can be delayed, hidden or transferred to a client.

The platform must therefore distinguish between work that can be automated, work that can be assisted, work that requires accountable professional judgement and work that should remain human.

In short:

• The acquisition thesis can span the physical and service economies.

• But the governance architecture cannot be identical across both.

The data flywheel is not automatic

Investors often describe data as though collecting more of it inevitably improves the model. But that’s not always the case.

For instance, a portfolio can accumulate vast quantities of inconsistent records, unlabelled files, duplicated documents, incorrect entries, biased decisions, expired procedures, customer-restricted data, personal information and machine telemetry without context.

The result will be a data swamp rather than a flywheel.

For the portfolio to learn, management needs common definitions, clear ownership, data lineage, access controls, retention rules, quality standards, customer permissions, model-evaluation methods and mechanisms for correcting errors.

It also needs evidence of which business outcomes improved.

The platform must understand where standardisation becomes destructive:

A common chart of accounts can be useful. A common safety assumption across different industrial processes may not be.

A shared customer-service model can improve consistency. It may also erase the local relationship that made the acquired business valuable.

The platform must know what should be common and what must remain specific.

That judgement will become one of its most important capabilities.

Productivity is not margin expansion

An acquired company can improve its margin without becoming more productive:

It can raise prices, reduce maintenance, delay investment, cut training, increase workloads, renegotiate suppliers aggressively, reduce service quality or use more debt.

Those actions may well raise short-term earnings but they do not necessarily increase useful output per unit of resource.

Likewise, AI can produce apparent savings while moving costs elsewhere:

A customer could wait longer to reach a person, an employee spend more time correcting automated work, a supplier be required to absorb additional administration, and a safety risk may grow but remain unrealised.

In short, a system can generate more output without increasing revenue or reducing total cost.

A credible transformation platform should therefore measure more than EBITDA.

It should examine output per hour, defects, downtime, lead times, energy and material consumption, customer retention, employee turnover, workplace injuries, working capital, innovation cycles, human correction, revenue from new products and the total cost of technology, integration and governance.

The platform succeeds when the business creates more verified value from its people, capital and physical resources, not merely when the reported margin improves before resale.

The seller’s decision changes

For the retiring owner, the choice may no longer be limited to family succession, a management buyout, employee ownership, a strategic sale, private equity or closure.

A new category could emerge:

That buyer would offer liquidity to the owner, a transition plan for management, technology investment, shared operating infrastructure, access to specialist capital, a route to combine with similar businesses and potential participation in future upside.

The offer could be attractive.

But it will also create old and new diligence questions for the seller:

• What will happen to the employees?

• Will the company retain its identity?

• Will the site remain open?

• Who owns the operating data?

• Can customer information be transferred into a shared model?

• How much debt will be placed on the business?

• What technology investment is contractually committed?

• Will management participate in equity?

• What happens if the transformation fails?

A seller who spent decades building a business is more likely than not to care about more than the highest headline price. The quality of the next owner will become part of the market.

The buyer’s advantage may be patience

A platform of this kind could fail if its capital expects rapid transformation:

Industrial processes do not change at software speed.

Equipment has replacement cycles.

Employees need training.

Customers require validation.

Regulators may need evidence.

Factories cannot be rebuilt every quarter.

Data may take years to organise.

And, operational improvements are likely to require investment before they produce savings.

The model could fail in precisely the unusual situations where experienced workers add most value.

For instance, a five-year fund structure, aggressive leverage and a near-term exit target could conflict with the time required to build a genuine intelligence platform. Therefore, patient capital is most likely to be an important competitive advantage.

The best owner may not be the one with the cheapest model, it is likely to be the one able to wait while the company learns how to use it.

The CFO version

For the CFO, the thesis is a capital-allocation problem.

The acquisition case should separate the value of the existing business from the expected operating improvement. It should identify the costs of technology, integration, workforce transition, data preparation, cyber controls, governance and physical automation.

It should also identify the risk that projected benefits do not materialise and the extent to which the financial model assumes that a future buyer will pay a higher multiple for an AI-enabled platform.

The CFO should resist the temptation to treat all digital expenditure as transformation investment. Some of it will be maintenance, experimentation, rework, vendor dependency, parallel running, failed deployment or necessary control infrastructure.

The investment committee should ask which improvement depends on AI, which depends on ordinary software, which depends on better management, which depends on scale and which depends on reducing employment and other synergies.

It should also ask what remains proprietary when the model becomes a commodity.

The financial model should not assume that adding AI raises the exit multiple: it should demonstrate how the business becomes economically better before assuming that a future buyer will pay more for it.

The CEO version

For the CEO of the acquired company, the challenge is sequencing.

The company cannot transform every process simultaneously. It should begin with workflows that are material, measurable, repetitive enough to improve, bounded enough to control, supported by usable evidence and connected to a clear business outcome.

It should avoid beginning with the most ‘theatrical’ application.

The useful first project may not be a humanoid robot walking through the factory. It may be predicting equipment failure, retrieving the correct engineering drawing, reducing quotation time, identifying quality drift, improving work scheduling, automating invoice matching or capturing knowledge before an expert retires.

The CEO should also decide what must not change: the customer relationship, safety culture, specialist craft, local decision right, exception process or identity of the business.

Transformation requires a positive operating thesis, not merely a list of old things to remove.

The employee version

For employees, the relevant promise is not that AI will create jobs somewhere in the economy.

It is what happens here.

To this job. To this site. To this pay packet. To this skill. To this pension. To this community.

Management should be able to explain which tasks are changing, which decisions remain human, what data are being collected, how performance will be assessed, which roles may disappear, which roles may grow and what training is available.

It should explain whether training occurs during paid time, how employees share in productivity gains, how an automated decision can be challenged and what happens if the technology fails.

Vague reassurance damages trust.

So does false certainty.

Some jobs will probably disappear even under a well-designed transformation.

Management should say that plainly.

The question is whether the process provides people with agency, evidence and a credible share of the opportunity.

The investor version

For the investor, the most important question is whether the platform has built a compounding asset or merely acquired a portfolio.

A real platform should become better at finding businesses, pricing transition risk, capturing knowledge, integrating data, deploying technology, training employees, measuring outcomes, controlling risk, developing management and transferring lessons.

A weak roll-up becomes more complicated with every acquisition.

In contrast, a strong platform becomes more capable.

The evidence should appear in declining integration time, improving operating outcomes and reduced dependence on external consultants.

The investor should also examine concentration. If every portfolio company depends on the same model, cloud provider, robotics stack, identity system, central data layer and integration team, efficiency may create systemic fragility.

The portfolio may look diversified by industry while sharing one operational point of failure.

The policymaker version

For policymakers, the Great Ownership Transfer is not merely a tax or retirement issue.

It touches employment, regional resilience, competition, productivity, access to capital, employee ownership, data rights, training and industrial strategy.

The closure of a small business can remove more than jobs. It can remove a service, local supplier, tax base and route into ownership.

Policy can reduce friction by supporting seller preparation, buyer education, management and employee buyouts, small-deal finance, portable training, standardised transaction information, regional succession networks, technical assistance and responsible data infrastructure.

However, policymakers should also examine what happens when thousands of local businesses are consolidated into technology platforms.

The economy could gain productivity while losing competition and local control, and a platform that rescues businesses from closure could become too powerful.

Succession policy and competition policy therefore cannot be separated indefinitely.

The insurance question

An AI-enabled acquisition platform creates risks at several levels:

• At the acquired company, a model can recommend an incorrect maintenance action.

• A robot could damage equipment or injure someone.

• An automated design may contain a defect and a professional-services system may produce incorrect advice.

• An employee may rely on an output beyond its approved purpose.

• Confidential information could be disclosed.

• A cyberattack could compromise connected operational systems.

At the portfolio level, one software error could then affect many companies with a shared model producing correlated mistakes. A central identity failure may interrupt operations across the group, a cloud outage may stop multiple businesses simultaneously and a compromised update may propagate through the platform.

At transaction level, diligence may miss a material liability, expected productivity gains fail, customer contracts restrict data use and historical records not support the intended models. Of course, the platform may inherit environmental, product or employment liabilities.

The insurance programme must therefore do more than add a generic AI endorsement.

It needs to map cyber, technology errors and omissions, professional indemnity, product liability, general liability, property damage, business interruption, crime, directors’ and officers’ liability, employment practices, environmental liability, representations and warranties, machinery breakdown and supply-chain exposure.

The boundaries will depend on the facts and policy wording.

An AI-related loss does not automatically fall within one category. A defective model-supported design may produce product liability, professional liability, property damage and business interruption across several customers.

In addition, a shared platform may also create accumulation. An insurer may believe it is underwriting dozens of independent small businesses when all depend on the same model and operating system.

The better the platform becomes at centralising intelligence, the more carefully it must manage correlated failure.

Insurability will depend on evidence. What was automated? What remained human? Which model and version were used? Which data informed the decision? Who approved the workflow? What authority did the system have? How was it tested? What happened when confidence was low? Could the action be stopped? Could the decision be reconstructed? How many portfolio companies depended on the same component?

The platform that can answer those questions will be easier to finance and insure than the one that merely says it uses AI.

The Board version

The board does not need to decide which robotics system should move a component, which model should analyse a maintenance record or which employee should build the first internal application.

It does need to decide what kind of owner the company intends to become.

An AI-enabled acquisition platform is not simply a technology programme attached to a roll-up. It changes the investment thesis, the operating model, the workforce bargain and the concentration of risk across the portfolio.

A useful board discussion should cover ten questions:

1. Are we acquiring fundamentally good businesses that lack successors, or using succession pressure to buy weak businesses cheaply?

2. Which part of the investment return comes from genuine productivity improvement, and which part depends on leverage, headcount reduction, price increases or a higher exit multiple?

3. What knowledge makes each acquired company valuable, where does that knowledge currently reside and how will it be preserved during the transition?

4. What should be standardised across the portfolio, and what must remain local because it reflects customer relationships, specialist judgement, safety requirements or operating context?

5. Which AI and automation investments have a defined business outcome, an accountable owner and a credible measure of value?

6. How will employees participate in designing the new operating model, and how will they share in the productivity gains they help create?

7. Which decisions can be delegated to machines, which require human approval and which should remain exclusively human?

8. How dependent will the portfolio become on common models, cloud providers, robotics systems, datasets, identity controls and central technology teams?

9. Can management reconstruct an AI-supported decision, stop a failing system and demonstrate where responsibility sits when something goes wrong?

10. Does the capital structure allow enough time to improve the businesses properly, or does the required return force management towards rapid extraction and resale?

The board should receive evidence across four dimensions.

1. The acquired businesses. Are revenue, customer retention, quality, output, safety, working capital and resilience improving?

2. The workforce. Are pay, skills, autonomy, job quality, mobility and employee participation improving, or is technology merely increasing work intensity?

3. The platform. Is each acquisition becoming easier to integrate because the system is learning, or is complexity accumulating faster than capability?

4. The risk. Are centralisation and common technology creating correlated exposures that the company, its insurers and its capital providers do not yet understand?

The board should be particularly cautious when management presents technology adoption as evidence of transformation. More models, robots, agents, dashboards and employee licences do not establish that the business is becoming more productive.

Nor should the board accept a simple choice between protecting employment and improving efficiency.

In the end, a business that fails to modernise is likely to protect neither.

But a company can modernise badly. It can destroy tacit knowledge, weaken customer relationships, impose unsuitable central systems, transfer risk to employees and record the resulting cost reduction as productivity.

The board therefore needs to distinguish between three things:

1. Automating work.

2. Improving the business.

3. Creating durable value.

They are not the same.

So, the hardest question is not whether the company has found enough businesses to acquire or enough AI to deploy.

It is:

The hard truth

The Great Ownership Transfer could become one of the largest economic-mobility opportunities of the next decade.

It could also become one of the largest transfers of locally accumulated knowledge into a small number of capital and technology platforms.

Both outcomes can arise from the same transaction.

A retiring owner sells. A business survives. Capital modernises it. Employees receive better tools. Productivity rises. Customers benefit. The company grows.

Or a retiring owner sells. Knowledge is extracted. Operations are centralised. Employment falls. Debt rises. Local control disappears. The business is resold.

The difference will not be determined by whether the buyer uses advanced AI.

It will be determined by who owns the platform, how it is financed, what it measures, how long it can wait, how it treats workers, how it shares gains, how it manages concentration and whether it builds companies or merely prepares them for resale.

Leave a comment

The weekend takeaway

The Great Ownership Transfer is usually presented as a race to find enough buyers.

That understates the opportunity.

The next generation of owners may inherit millions of functioning businesses at the same moment that artificial intelligence becomes capable of changing how those businesses design, manufacture, sell, maintain and learn.

Jeff Bezos may be assembling an unusually powerful version of that model: a large pool of acquisition capital, an AI laboratory focused on engineering and physical production, and experience operating robotics and AI across one of the world’s largest physical networks.

If those pieces come together, they could create a platform for industrial productivity at exceptional scale.

The result is not guaranteed.

The proposed investment vehicle may not reach its reported target. Prometheus may not achieve its technical ambitions. Technologies developed for one operation may not transfer easily into thousands of different businesses. The employment gains predicted by Bezos may not arrive evenly or quickly.

But the underlying possibility remains.

The most valuable AI company may not be the company that owns the best chatbot.

It may be the company that learns how to transfer, preserve and improve the productive knowledge of an entire generation of businesses.

The owners are retiring.

The companies are not.

The question is who acquires them, what intelligence they install and who participates in the value created when the machines begin to learn.

🧠 Techie Corner for the Non-Techies

1. The Great Ownership Transfer

The term describes the expected transfer of large numbers of privately owned businesses (by Baby Boomers) as their owners retire.

2. Entrepreneurship through acquisition

Entrepreneurship through acquisition means buying and operating an existing business rather than starting one from nothing.

The buyer acquires existing customers, employees, systems and cash flow. The model can reduce some start-up risks but creates others, including financing, succession and integration risk.

3. Roll-up

A roll-up combines multiple businesses in the same or adjacent sectors under one ownership structure. The aim is to create scale, centralise functions, improve purchasing, cross-sell services or increase the combined company’s valuation.

Roll-ups can also create excessive leverage, integration problems and market concentration.

4. Platform company

In private equity, a platform company is often the initial or principal business around which further acquisitions are made.

In this article, platform has a broader meaning: a common operating, data and technology system that can improve multiple acquired businesses.

Owning several companies does not automatically create a platform.

5. Physical AI

Physical AI refers broadly to AI systems that perceive, model or act within the physical world.

Applications can include robotics, manufacturing, autonomous vehicles, industrial simulation, engineering design, warehouse automation and predictive maintenance.

The term is used inconsistently and does not denote one settled technical category.

6. Artificial general engineer

Prometheus uses this phrase for its ambition to build AI systems capable of assisting across the engineering process, including design, simulation and manufacturing.

It should not be confused with a confirmed achievement of artificial general intelligence.

7. Foundation model

A foundation model is trained on broad data and can be adapted to many tasks.

Large language models are one type. Models for images, robotics, biological systems or physical simulation can use different architectures and data.

Access to a foundation model does not by itself create proprietary advantage.

8. World model

A world model attempts to represent aspects of how an environment behaves.

In physical applications, this often includes predictions about movement, force, materials or cause and effect.

The term covers different technical approaches and should not be taken to mean that a system understands the physical world in the same way a person does.

9. Tacit knowledge

Tacit knowledge is practical understanding that is difficult to express fully in formal documents.

It includes recognising an unusual sound, knowing when a customer is likely to reject a proposed solution, understanding which supplier can deliver during an emergency or sensing when a machine is operating abnormally.

Capturing tacit knowledge is difficult because asking someone to describe it may change or simplify what they actually do.

10. Knowledge graph

A knowledge graph represents relationships between entities such as customers, products, people, contracts, machines and events.

It can help an AI system retrieve connected information.

A knowledge graph is only as reliable as its data and definitions.

11. Citizen developer

A citizen developer is an employee who creates applications or workflows without being a professional software developer.

AI-assisted coding makes this easier, but it also increases the need for security, permissions, testing and governance.

12. Shadow AI

Shadow AI is the use of AI tools outside approved organisational controls.

Examples include uploading confidential documents to a personal chatbot or deploying an unapproved agent.

The term resembles shadow IT, where employees adopt technology without formal authorisation.

13. Model routing

Model routing assigns different tasks to different models or systems.

A simple task may use ordinary software or a smaller model. A complex engineering problem may require a more capable system.

Routing can reduce cost and risk when it is designed properly.

14. Human in the loop

A human-in-the-loop system requires a person to review or approve part of the AI-supported process.

The phrase can conceal weak control. A person who lacks time, information or authority could provide only superficial oversight.

The relevant question is where human judgement enters and whether it can alter the outcome.

15. AI-enabled value creation

AI-enabled value creation means using AI to improve an economic outcome such as revenue, cost, quality, speed, safety or risk.

It should be distinguished from AI activity: more prompts, models or generated documents do not prove that value has been created.

16. Learning arbitrage

Learning arbitrage is used in this article to describe value created when lessons from one business improve the operation or integration of another.

The value depends on whether knowledge can genuinely be transferred without losing important context.

17. Data flywheel

A data flywheel is the idea that more use generates more data, which improves the system, which attracts more use.

The cycle is not automatic: poor-quality or biased data can degrade a system rather than improve it.

18. Operational technology

Operational technology controls or monitors physical equipment and industrial processes.

It differs from conventional information technology, although the two increasingly connect.

A cyber incident affecting operational technology can create physical damage and have safety consequences.

19. AI accumulation risk

AI accumulation risk occurs when many companies depend on the same model, cloud provider, dataset or software component.

One failure may then produce losses across several organisations and insurance policies simultaneously.

20. Employee ownership

Employee ownership gives employees an economic interest in the company.

Structures vary by jurisdiction and can include employee share plans, trusts, cooperatives and direct ownership.

Employee ownership does not automatically produce employee control or equal participation.

📚 Further reading

• McKinsey Institute for Economic Mobility: The Great Ownership Transfer: A New Era of Business Stewardship(2026)

  The report examines the expected transfer of US small and medium-sized businesses as owners retire. It considers regional exposure, transaction barriers, economic mobility and the need for more buyers.

• Ken Yearwood: A Generational Shift in Business Ownership Is Underway (2026)

  A shorter McKinsey commentary on succession risk, business continuity and the opportunity to broaden access to business ownership. It argues that delayed planning can place companies, jobs and community wealth at risk.

• The Wall Street Journal: Jeff Bezos in Talks to Raise $100 Billion for AI Manufacturing Fund (2026)

  The report states that Bezos was in early discussions to raise a manufacturing transformation vehicle intended to acquire industrial companies and deploy AI across them.

• The Wall Street Journal: Bezos Bats Down AI Job Loss Fears While Launching New Venture (2026)

  An interview and report on Prometheus, its funding, engineering ambitions and Bezos’s view that AI may create labour shortages.

• Reuters: AI Will Lead to Labour Shortages, Bezos Says in Optimistic Talk (2026)

  Reuters’ report from VivaTech on Bezos’s employment thesis and the stated purpose of Prometheus. It also notes that companies are currently attributing some job reductions to AI-related efficiency.

• Project Prometheus: Company launch materials (2026)

  The company describes its ambition to build an artificial general engineer for designing and manufacturing physical products.

• Amazon: Amazon Deploys Over One Million Robots and Launches New AI Foundation Model (2025)

  Amazon reports that it had deployed one million robots across more than 300 facilities and introduced DeepFleet to improve robotic-fleet movement.

• Amazon: Amazon Has Upskilled Over 700,000 Employees Globally (2025)

  Amazon’s description of its education, apprenticeship and workforce-development programmes.

• Amazon: Amazon unveils next-gen Proteus robot as part of €10 billion European investment in its fulfillment network (2026)

  A company announcement covering newer robotics systems, European investment and workforce plans.

• IT Brew: How Walmart Is Securing a New Cohort of AI Builders (2026)

  An account of Walmart’s Code Puppy platform, non-technical employee development and the company’s approach to permissions and embedded governance.

• USAFacts: How Many Americans Are Using AI- and How? (2026)

  A summary of experimental US Census Bureau survey data collected in June 2025. The data show household-reported AI use, productivity perceptions, trust and demographic differences. The survey measures whether respondents reported AI use within their household. It is not a direct count of unique individual users or audited workplace deployment.

• Markets Group and Brownloop: The AI Illusion in Private Equity: Why Using AI Is Not the Same as Owning It (2026)

  A sponsored practitioner article arguing that generic AI adoption does not create durable investment advantage without proprietary data, institutional knowledge and governance.

• Dr Henri Winand: Rent the Intelligence. Own the Proof. (2026)

  An earlier Something for the Weekend article on the distinction between access to models and ownership of evidence, workflow and institutional knowledge.

• Dr Henri Winand: The Greenest AI Is the One That Wastes Least (2026)

  The article argued that AI efficiency should be measured as verified business value per unit of total resource consumed.

Last week on Something for the Weekend

Last week’s piece asked whether companies were becoming more efficient or merely doing more machine work.

It argued that falling model costs, lower energy per task and rising usage can occur simultaneously.

The correct corporate measure is not machine output alone, it is verified business value per unit of total resource consumed.

This week moves from the AI workflow to the company itself.

As millions of businesses approach succession, the opportunity may be to connect established customers, skilled workers and accumulated knowledge with a new generation of intelligent tools.

The machines may improve, and the ownership model will decide who benefits.

The board receives its quarterly AI report:

Usage is up.

More employees have access to copilots.

Several hundred agents have been launched.

The cost of some models has fallen.

The average response arrives faster.

The sustainability team reports that the estimated carbon intensity of each request is improving.

It sounds encouraging, then the CFO asks three questions:

1. How much of the old work disappeared?

2. How much verified value reached the profit and loss account?

3. Why did the total AI bill still rise?

The CIO explains that the company is doing much more with AI. More users. More prompts. More documents. More code. More agents. More automated decisions.

That may be true.

However, more machine activity is not the same as more business value.

The company may have made each unit of AI cheaper while buying many more units. It may have reduced the energy required for a simple request while replacing simple requests with agents that read files, call tools, reconsider their plans, retry failures and ask other models to check their work.

The environmental report may also be less certain than it appears.

For most closed AI services, the corporate user cannot see the physical machine answering the request. It cannot see the precise hardware, the number of other users sharing it, the utilisation rate, the cooling system, the electricity source or the location from which the workload was served.

The carbon number is therefore often an estimate built from assumptions and a precise decimal point does not remove that uncertainty.

The board is left with a more useful question:

The Boardroom point

AI efficiency should not be measured by how much output the machine produces. It should be measured by the verified business value produced for every unit of total resource consumed.

That includes:

Tokens.

Compute.

Electricity.

Water.

Carbon.

Cloud expenditure.

Hardware.

Human verification.

Correction and rework.

Management attention.

Operational risk.

The numerator matters as much as the denominator:

• A model that consumes little energy but produces a wrong answer is not efficient.

• A cheap agent that saves no human time is not productive.

• A fast system that creates work nobody uses has not created value.

• A smaller model that repeatedly fails may be more expensive than a larger model that completes the task correctly.

The useful measure is neither output per kilowatt nor it is not tokens per pound. It is not documents per employee either.

It is:

That is where the environmental argument and the commercial argument begin to converge: waste is expensive before anyone calls it unsustainable.

Last week, the token became a kilowatt

Last week’s Something for the Weekend followed the AI token down through the software stack:

The token looked weightless.

Underneath it sat inference.

Underneath inference sat chips, memory, networks and cooling.

Underneath those sat data centres, grid connections, power contracts, land, capital and water.

The conclusion was that the token is only the basic meter.

But a better meter is useful, verified work.

This week’s question is what happens when companies install millions of those meters without first deciding which machine work is worth paying for.

That is important because AI is becoming more efficient at the same time as it is becoming much more widely used. Those two statements are not contradictory: a more efficient engine can still consume more fuel if the company installs thousands of engines and runs them for longer.

The International Energy Agency reported in 2026 that global data-centre electricity demand increased by 17% during 2025. It estimated that consumption in AI-focused data centres increased by about 50% over the same year.

These are system-level estimates, not direct measurements of every data centre. They should therefore be treated as modelled aggregates rather than perfect counts.

But the direction is nevertheless clear:

• The machines are becoming more efficient.

• The demand for machine work is growing faster.

The IEA’s central projection has global data-centre electricity consumption rising from approximately 485 terawatt-hours in 2025 to around 950 terawatt-hours in 2030. It expects electricity consumption from AI-focused data centres to grow much faster than the broader category, roughly tripling over that period.

Terawatt-hours can be difficult to picture. So, put into terms we all see every day, the electricity consumed by the world’s data centres in 2025 was equivalent to the annual electricity use of approximately:

• 186 million typical Chinese homes;

• 180 million typical British homes;

• 143 million typical German homes;

• 101 million typical Japanese homes; or

• 47 million typical US homes.

These are alternative ways of describing the same 485 terawatt-hours. They are not numbers to be added together. By 2030, the IEA’s central projection would be equivalent to the annual electricity consumption of approximately:

• 364 million Chinese homes;

• 350 million British homes;

• 280 million German homes;

• 198 million Japanese homes; or

• 92 million US homes.

The numbers differ because a typical household does not consume the same amount of electricity in every country. Climate, home size, electric heating, air-conditioning, appliances and household size all matter.

But whichever domestic benchmark is used, the scale is difficult to ignore.

The world is moving from enough data-centre electricity to power tens or hundreds of millions of homes towards nearly twice as much within five years. Whilst that projection may change, AI demand, hardware efficiency, power availability, financing conditions and regulation are all moving quickly.

However, the boardroom implication does not depend on the exact forecast being correct:

Efficiency per task will not control the company’s total AI footprint if the number and complexity of tasks grow faster.

The digital co-worker is not a unit

Companies increasingly refer to AI agents as digital employees or digital co-workers.

Whilst that phrase is useful, it is also dangerous:

• There is no standard unit called one digital co-worker.

• One may answer short questions from employees.

• Another may review contracts.

• Another may watch transactions for fraud.

• Another may search a software repository, modify code, run tests, inspect failures, rewrite the code and try again.

• Another may communicate with customers, retrieve account information, calculate an answer, trigger a refund and create a record for human review.

Calling all of them digital workers is like calling a bicycle, a delivery van and a cargo aircraft transport. The label describes their purpose but it does not describe what they consume. The energy footprint of a digital co-worker depends on:

The model used.

The length of its instructions.

The amount of information it reads.

The number of steps it takes.

The tools it calls.

The number of times it fails.

The number of times it retries.

The length of its output.

The hardware on which it runs.

The utilisation of that hardware.

The data-centre overhead.

The electricity mix.

The human work needed to check the result.

Without those facts, an average annual carbon footprint for a digital employee would create a false sense of precision.

A corporate user should not ask:

What is the footprint of one AI employee?

It should ask:

What does this particular AI workflow consume to produce one accepted business outcome?

That outcome might be:

• A correctly resolved customer enquiry.

• A software defect repaired and passed through testing.

• A contract reviewed to an agreed standard.

• An invoice matched without later correction.

• A claim triaged and accepted by a human reviewer.

• A board paper produced with verified evidence.

• A transaction flagged and correctly investigated.

That is a functional unit which makes comparison possible because it connects the machine to the job.

The agent can spend without seeing the bill

A human employee has an employment contract:

• They have working hours.

• They have delegated authority.

• They may have a company credit card and an expense limit.

• They usually know when they have spent a day on a task.

An AI agent operates differently:

• It can work continuously.

• It can call other services.

• It can load more information.

• It can ask another model for help.

• It can abandon one route and try another.

• It can repeat the same mistake at machine speed.

• Most importantly, it does not feel the cost.

Recent academic work illustrates the problem, although the findings should not yet be treated as universal. A 2026 preprint titled How Do AI Agents Spend Your Money? studied eight frontier models performing agentic software-engineering tasks. The researchers reported that these tasks consumed around 1,000 times more tokens than the code-chat and code-reasoning comparisons they used.

Runs on the same task differed by as much as 30 times in total token consumption.

More token consumption did not consistently produce higher accuracy. In the study, performance often peaked at an intermediate level of expenditure and then stopped improving materially.

The researchers also found that the tested models were poor at forecasting their own token requirements and tended to underestimate them.

These findings come from coding tasks and may not transfer directly to law, finance, insurance, manufacturing or customer service. The paper was also a preprint at the time of writing, not a settled industry standard.

The management lesson is still important:

1. An agent may not know how much work it will consume.

2. More machine effort may not produce a better result.

3. The same task may produce materially different costs on different runs.

Where the waste hides

The visible AI invoice is only part of the cost.

Waste hides throughout the workflow.

The model is too large for the job

A frontier model may be justified for difficult research, complex reasoning or broad synthesis. Yet, it may be unnecessary for classifying documents, matching invoices, extracting defined fields or checking whether a form is complete.

Using the largest available model for every task is not a sign of technological ambition:

It is a failure to route work properly.

It is the digital equivalent of sending a long-haul aircraft to deliver an envelope.

The machine reads too much

Large context windows allow AI systems to process long documents, code repositories and conversation histories. That does not mean the company should give every model every available document.

Much of the information may be irrelevant and the model still has to process it. Better systems retrieve the relevant material first and then ask the model to reason from the smaller set.

The objective is not the smallest possible prompt. Over-compressing information can remove facts the model needs and produce worse answers.

The objective is the smallest sufficient body of evidence for a reliable result.

The agent takes the scenic route

An agent can search, inspect, plan, revise and retry. That flexibility is part of its value.

It is also a source of waste:

• An agent without a step limit may continue exploring after the economically useful point has passed.

• An agent without a retry budget may repeat the same failure.

• An agent without an escalation rule may spend more trying to solve an exception than the exception is worth.

A scenic route can be enjoyable on holiday, it is less attractive when the taxi meter is running.

The machine talks too much

Longer answers consume more output tokens and generally require more generation work.

Some tasks need detail. Many do not.

A 2026 preprint examining open-weight language models in software-development tasks found that several models produced unnecessarily verbose output. The researchers reported large energy savings after suppressing this behaviour in their tested code-generation settings without reducing measured accuracy.

The reported savings ranged from 44% to 89%.

That is a striking result, but it should not be generalised casually. The study examined a limited set of models and coding benchmarks under controlled conditions. It does not prove that every company can cut its AI energy use by the same amount.

It does demonstrate a simpler point: the machine can consume resources by saying more than the job requires.

The answer is wrong

A wrong answer consumes resources when it is produced:

It consumes more when someone checks it.

It consumes more again when the workflow is corrected.

A hallucination is not only a quality problem, it is waste with a confidence problem.

The model may consume tokens to produce the answer. A lawyer, engineer, accountant or underwriter may then consume expensive human time finding the error. The system may be called again to repair it.

Yet, the company pays for the generation and the correction. That is verification debt with an energy bill attached.

Nobody uses the output

AI makes generation cheap. It can produce more reports, summaries, options, presentations, images and code than the organisation can absorb. Output that is never read can still appear in a productivity dashboard.

It may even look impressive:

The machine was busy.

The employees were experimenting.

The token count rose.

Nothing changed.

A digital worker can look productive while creating no more value than a human worker sending emails to appear occupied. It can simply do it faster with a meter attached to it.

The old process remains

This is one of the most common ways for an AI business case to fail:

• The company introduces an AI step.

• It keeps the old step.

• It adds human review.

• It adds monitoring.

• It adds governance.

• It adds another vendor.

The process becomes faster at producing an intermediate output but more expensive overall.

If the old work does not disappear, the company may have added a layer rather than removed one.

Leave a comment

The greenest answer may not be the shortest

There is a temptation to make the argument too simple:

Fewer tokens must mean less energy.

Smaller models must be greener.

Shorter prompts must be better.

That is not always true:

A smaller model may fail repeatedly.

A compressed prompt may omit the facts needed for the answer.

A cheap model may create more verification work.

A short answer may be useless.

A local machine may sit underused.

A more capable model may complete a difficult task correctly in one attempt.

Recent research reinforces the need for caution. A 2026 preprint examining prompt compression across several model providers found that shortening the input did not reliably reduce estimated energy while preserving quality. In some tested settings, compression degraded task performance severely or caused models to produce much longer outputs.

The study estimated energy using a token-based proxy calibrated against local measurements. It did not directly measure the proprietary providers’ physical infrastructure. Its conclusion was therefore not that compression is bad. It was that reducing input tokens alone is not a reliable optimisation strategy.

That is the right corporate principle: do not optimise a visible metric while damaging the outcome that metric is meant to serve.

Thus, the greenest answer is not necessarily the shortest answer. It is the answer that achieves the required result with the least avoidable total consumption.

A carbon number is a model too

The same AI request can receive different carbon estimates from different methodologies. This is not surprising.

The methodologies may assume different:

Hardware.

Model sizes.

Batching levels.

Server utilisation.

Generation times.

Data-centre overheads.

Electricity sources.

Locations.

Cooling systems.

Hardware lifetimes.

Lifecycle boundaries.

Some estimates cover operational electricity only.

Others try to include emissions associated with manufacturing equipment.

Some use measured telemetry.

Others infer consumption from tokens, latency and model identity.

For a closed API, the external user usually cannot observe the whole physical system. That means environmental reporting may contain three different types of information:

1. Measured data.

2. Provider-supplied data.

3. Modelled estimates.

They should not be presented as though they carry the same level of certainty but this does not make estimation useless. Companies routinely use models where direct observation is difficult. The problem is false precision.

A credible report should say:

• What was measured.

• What was estimated.

• Which methodology was used.

• Which boundaries were included.

• Which assumptions were made.

• How uncertain the result may be.

• Whether two compared figures were calculated on the same basis.

The board should be suspicious of an environmental number that contains several decimal places but no methodology.

Measurement is beginning to catch up

The field is not starting from nothing. The Green Software Foundation ratified its SCI for AI specification in late 2025.

It extends the Software Carbon Intensity approach across AI systems, including generative and agentic workloads. It distinguishes between impacts a provider can influence and the operational impacts a consumer can influence.

The specification supports different functional units for different types of AI system:

A language model may be measured per token.

A classifier may be measured per inference.

A corporate workflow should ultimately be measured per useful result.

The Green Software Foundation has also argued that many organisations deploying agents still lack task-level tooling for energy, carbon and water. That assessment comes from an industry standards organisation rather than a comprehensive global audit, so it should not be read as a measured percentage of all companies.

It is nevertheless consistent with the practical problem faced by corporate users:

They can often see the cloud bill.

They may see token consumption.

They rarely see the full physical resources consumed by an individual business outcome.

MLCommons is developing and operating benchmark frameworks that measure performance and power together across machine-learning systems. These benchmarks improve comparison at the infrastructure level. They do not by themselves tell a board whether an AI-assisted underwriting decision, contract review or customer interaction created value.

Finance is discovering the same problem

Environmental measurement is not developing in isolation and financial management is moving in the same direction.

In June 2026, the Linux Foundation announced its intention to establish a Tokenomics Foundation focused on open standards, benchmarks and practices for AI infrastructure economics. The announcement described a planned collaboration with the FinOps Foundation and support from organisations including major technology providers and enterprise buyers.

It was an announcement of intent, not proof that a mature standard had already been delivered. Its timing is still revealing.

Enterprise AI has moved far enough into production that buyers now need a common language for comparing models, providers, tokens and outcomes.

The FinOps Foundation’s 2026 guidance makes the same operational point: AI costs do not sit only in model inference, they appear across training, tuning, retrieval, orchestration, monitoring, security, subscriptions, cloud infrastructure and agent workflows.

The FinOps guidance warns that poorly controlled agents can create compounding expenditure through repeated calls and loops. That wording should not be interpreted mathematically in every case. Not every agent bill grows exponentially.

The useful observation is that one action can trigger several more actions, each carrying its own cost. Therefore, AI spend can multiply through the architecture of the workflow.

The finance department cannot manage this by counting software seats: it needs to understand the use case.

Green versus fierce is a false choice

The debate is often framed badly:

On one side are people concerned about emissions, water and energy consumption.

On the other are people who want growth, productivity and industrial competitiveness.

The argument then becomes hugging trees versus drilling wells which misses the commercial point:

• A manufacturer does not reduce scrap because it dislikes production, it reduces scrap because wasted material destroys margin.

• An airline does not reduce unnecessary weight because it opposes travel, it does so because fuel has a cost.

• A retailer does not improve inventory turns because warehouses are immoral, it does so because idle stock consumes capital.

AI is no different:

• Unnecessary tokens consume money.

• Unnecessary context consumes compute.

• Unnecessary retries consume capacity.

• Unnecessary output consumes review time.

• Wrong answers consume trust.

The efficient AI system is not less ambitious, it is better designed.

The fiercest competitor will not necessarily use the least AI:

It will use the least wasteful AI.

It will know which work requires a frontier model.

Which work belongs with a smaller model.

Which work should run locally.

Which work needs retrieval.

Which work should be handled by ordinary software.

Which work requires a person.

Which work should not happen at all.

That is not environmental restraint imposed on corporate growth, it is operating discipline.

The rebound problem

There is an awkward complication: efficiency often increases total consumption.

The economist William Stanley Jevons observed in the nineteenth century that improvements in the efficiency of coal use could increase rather than reduce total coal consumption by making more uses economically attractive.

The modern application is sometimes called the Jevons paradox or rebound effect. The analogy should not be treated as a mechanical law.

Cheaper AI does not guarantee that every company will consume more of it, but it makes more machine work economically possible:

As cost per token falls, companies can automate tasks that were previously uneconomic.

As models improve, employees find new uses.

As agents become capable of longer workflows, one human request can trigger many machine actions.

The energy consumed by each action may fall while the total number of actions rises much faster.

This is why a board needs two sets of measures.

First, intensity:

• Cost per verified outcome.

• Energy per verified outcome.

• Carbon per verified outcome.

• Human review time per verified outcome.

Second, absolute consumption:

• Total AI expenditure.

• Total tokens.

• Total electricity attributable to the system, where it can be measured or reasonably estimated.

• Total estimated emissions.

• Total human verification time.

A company can halve the energy used per transaction and increase the number of AI-assisted transactions tenfold. So, both measures must be considered in concert: an improving ratio can coexist with a rapidly expanding total footprint.

The CFO version

For the CFO, the problem is no longer simply the AI budget, it is the economics of machine work.

The finance team needs to ask:

• What is the cost per completed outcome?

• What is the cost per accepted outcome?

• What is the difference between the two?

• How much human review is required?

• What does a failed run cost?

• Which business unit created the consumption?

• Which model and provider were used?

• Which workflows generate revenue?

• Which workflows reduce cost?

• Which workflows reduce risk?

• Which old processes have been removed?

• Which costs remain hidden in cloud, software, consulting, legal, compliance and security budgets?

The fully loaded cost of AI includes more than the provider’s invoice, it includes integration:

• Data preparation.

• Retrieval.

• Monitoring.

• Security.

• Human checking.

• Failure recovery.

• Governance.

• Vendor dependency.

• The cost of keeping the old process alive.

A falling cost per token may be good news, but it is not evidence of a rising return on investment.

The question is: which AI consumption reaches the P&L as value?

The CIO version

For the CIO, this is an architecture problem. The enterprise needs routing rules:

• Which tasks need broad reasoning?

• Which need a narrow model?

• Which require local control?

• Which can use a public cloud service?

• Which should use deterministic software?

• Which require a source record?

• Which require human approval?

• Which can be cached?

• Which contexts can be shortened safely?

• Which agents need step limits?

• Which agents need spending limits?

• Which failures should trigger another attempt?

• Which failures should trigger escalation?

• Which workflows should be terminated?

The objective is not to minimise tokens blindly, it is to reduce avoidable waste without losing the quality, security or accountability the business requires:

Use enough model for the task.

Use enough context for the answer.

Use enough verification for the risk.

Use enough evidence for the decision.

Use enough human judgement for accountability.

No more, no less.

The sustainability version

For the sustainability team, the problem is methodological credibility. A responsible AI footprint report should distinguish between:

• Direct measurement.

• Provider disclosure.

• Third-party estimation.

• Internal modelling.

It should state the boundary.

• Does the number cover only inference electricity?

• Does it include data-centre overhead?

• Does it include cooling?

• Does it include hardware manufacturing?

• Does it include training?

• Does it include storage and networking?

• Does it cover water withdrawal, water consumption or neither?

• What electricity mix was assumed?

• Was the actual serving region known?

• Was an annual grid average used or a time-specific figure?

These questions do not make reporting impossible, they make honest reporting possible. In April 2026, the European Commission opened a consultation on measuring the energy consumption and emissions of AI models and systems.

The consultation was intended to inform a measurement framework connected to the AI Act’s energy-related objectives and the possible design of an AI energy and emissions label. The Commission noted that a comprehensive assessment requires information across computational resources, electricity use and hardware.

The consultation demonstrates that policymakers see the issue but it also demonstrates that the methodology is not finished.

Companies should prepare for greater disclosure without pretending that every current estimate is audit-grade.

The Board version

The board does not need to select the best GPU and it does not need to debate every prompt. However, it does need to understand whether management is building a machine-work cost base the company can control.

A useful board discussion should cover ten questions:

1. Where is AI being used in production?

2. Which workflows consume the most tokens, compute and human review?

3. Which AI costs are fixed, variable or hidden inside other budgets?

4. Which workflows have a defined unit of useful work?

5. Which agents have budgets, retry limits and stop conditions?

6. Which tasks are routed to unnecessarily large models?

7. Which AI outputs require verification?

8. Which old processes have actually been retired?

9. How are environmental estimates calculated, and what uncertainty remains?

10. What happens to cost and total resource consumption when usage increases tenfold?

The board should then ask the harder question:

The insurance question

Waste can become loss and an uncontrolled agent may consume more than expected:

It may access the wrong data.

It may trigger unauthorised actions.

It may generate advice on which someone relies.

It may disclose confidential information.

It may continue operating after the economically useful point has passed.

It may create operational, legal, cyber, professional-liability or reputational consequences.

The resulting loss may not fit neatly into one insurance policy:

• Was it a cyber event?

• A technology error?

• A professional-services failure?

• An employee-like act?

• An unauthorised transaction?

• A failure of internal controls?

• A product-liability event?

• Or several of them at once?

That uncertainty is relevant because insurers do not only ask how likely a loss is. They also need to know where the loss sits, how large it could become, whether several policyholders could suffer the same failure at the same time and which policy is supposed to respond.

AI makes each of those questions harder. The same foundation model may sit beneath thousands of companies. The same vendor may provide models, infrastructure, data services and security controls.

Therefore, one upstream failure could create losses across many insureds, industries and policy classes at the same time. That is an accumulation problem.

It helps explain why the insurance market is moving away from leaving AI exposures silently embedded inside conventional policies. Some insurers are introducing exclusions, clarifications or sublimits, while a smaller affirmative market is beginning to offer more explicit protection for defined AI risks.

The direction is not yet uniform. Coverage will depend on the policy wording, jurisdiction, loss mechanism and facts of the claim. A reference to AI in a loss does not automatically mean that every policy will exclude it. Nor does the absence of an AI exclusion guarantee that the claim will be covered.

But the broad lesson is clear:

An AI system with poor resource discipline may therefore be difficult to underwrite. An insurer examining a material AI workflow will increasingly want to know:

• What authority the agent has.

• Which data it can access.

• Which systems and tools it can operate.

• What financial, operational and time limits apply.

• Which actions require human approval.

• How its actions are recorded.

• How often humans intervene.

• How failures are detected.

• Whether the system can be stopped.

• Whether its decisions can be reconstructed.

• Which external model, cloud and data providers it depends on.

• What happens if one of those providers changes, fails or is compromised.

Yet, an efficient and well-controlled AI system is not automatically safe or insurable.

Controls do not remove the risk, they make the risk more legible. So, insurance becomes more plausible when the company can convert an open-ended AI exposure into a defined operating risk. That requires more than a general AI policy written by the legal department.

The insurer needs to see a bounded use case. The agent should have a stated purpose, a defined population of users, approved data sources and a clear list of permitted actions. It needs limited authority.

An agent that can draft a payment instruction presents a different exposure from an agent that can approve and transmit the payment.

An agent that recommends a customer response differs from one that sends it without review:

• It needs financial and operational limits. That may include spending caps, transaction limits, restricted counterparties, maximum run times, limits on tool calls, retry budgets and automatic escalation when the system moves outside its expected behaviour.

• It needs human control at the consequential point. Not every AI action requires approval. But decisions that move money, alter legal rights, affect safety, bind the company or materially affect a customer should have an identified accountable person or a deliberately approved exception.

• It needs evidence. The company should be able to show which model ran, which version was used, what data the system accessed, what instructions applied, which tools were called, what the system produced, who reviewed it and what final action followed.

• It needs testing against failure, not merely testing for average performance. That includes hallucination, prompt injection, data leakage, model drift, inappropriate tool use, vendor outage, corrupted retrieval, excessive autonomy and failure of the human escalation route.

• It needs a stop mechanism. A kill switch sounds dramatic until a system starts taking repeated actions at machine speed.

• It needs an incident plan. The organisation should know who investigates, who contains the event, who preserves the evidence, who informs customers or regulators and when the insurer must be notified.

• It also needs an insurance map. Cyber, technology errors and omissions, professional indemnity, crime, media liability, product liability, directors’ and officers’ liability and general liability policies may each respond differently.

The company should not wait for a loss to discover that every insurer assumed another policy was covering the exposure. That may lead to:

• affirmative AI coverage for a tightly defined risk;

• an endorsement under an existing policy;

• a dedicated sublimit, retention or aggregate;

• exclusions for activities the insurer will not accept.

It may require several policies to operate together. The objective is not necessarily to find one policy called “AI insurance”.

The more realistic answer may be a coordinated programme in which the different loss pathways are expressly allocated rather than left silent. That is what could make corporate AI use more insurable.

Evidence that the company knows what the system is allowed to do, how it can fail, can see what it actually did, can limit the consequences when it goes wrong and has agreed in advance where the resulting loss should sit.

The insurable AI system will not be the one described as intelligent, it will be the one whose risk can be bounded, observed, reconstructed and priced.

The weekend takeaway

The digital co-worker does not need a reusable coffee cup, it needs a properly designed job.

The wrong model:

• reads too much;

• thinks too long;

• retries too often;

• calls too many tools;

• produces more than the company can use; then

• asks a human to work out whether any of it was right.

The right model:

• does enough work;

• uses the evidence it needs;

• reaches a defensible answer;

• creates a record; then

• stops.

The environmental argument and the commercial argument are beginning to converge:

Waste is expensive in pounds.

Waste is expensive in kilowatts.

Waste is expensive in water, carbon and human attention.

The company with the most AI may look advanced.

🧠 Techie Corner for the Non-Techies

1. Token

A token is a unit of text or data processed by a language model. Tokens are not the same as words. A word may contain one token or several, depending on the model and the text.

AI providers often use input and output tokens as part of their pricing and usage systems. A token is a useful measure of model activity.

It is not a direct, universal measure of energy because different models, hardware and serving systems can consume different amounts of energy for the same number of tokens.

2. Inference

Inference is the process of using a trained AI model to produce an answer, classification, prediction or action.

Training builds or adjusts the model. Inference uses it.

Most corporate users will not train a frontier model from the beginning. They will consume inference directly through APIs, software products, cloud platforms or internal deployments.

3. Prefill

Prefill is the stage during which a language model processes the input.

That input may include:

• The user’s question.

• System instructions.

• Documents.

• Conversation history.

• Retrieved evidence.

The model converts that information into internal representations used to generate the answer. Long or unnecessary context can increase prefill work.

4. Decode

Decode is the stage during which the model generates output tokens. Longer answers generally require more decoding work.

The energy relationship is not perfectly linear across all models and hardware, but output length is an important driver of inference cost.

5. Context window

The context window is the amount of information a model can consider during one interaction. A larger context window allows the model to read more text, code or history.

It does not guarantee a better answer. Large contexts can increase cost, latency and energy use. They may also make it harder for the model to identify the relevant material.

6. Retrieval-augmented generation

Retrieval-augmented generation, usually called RAG, is a method that searches for relevant information before asking a language model to answer. Instead of placing an entire document collection into the prompt, the system retrieves a smaller set of relevant passages.

A well-designed retrieval system can reduce unnecessary context and connect an answer to evidence. A poorly designed one can retrieve the wrong evidence and create false confidence.

7. AI agent

An AI agent is a system that can take several steps towards a goal.

It may:

• Read documents.

• Search databases.

• Call software tools.

• Write code.

• Run tests.

• Revise its plan.

• Retry failed actions.

• Escalate to a human.

The agent is not necessarily autonomous in the human sense. Its authority depends on the systems and permissions given to it.

8. Model routing

Model routing sends different tasks to different models or systems:

• A simple task may use a small model.

• A complex task may use a frontier model.

• A rules-based task may not need a language model at all.

Routing attempts to match the capability and cost of the machine to the needs of the task.

9. Functional unit

A functional unit is the thing being measured:

• “Per token” is one functional unit.

• “Per inference” is another.

For a business, stronger functional units may include:

• Per resolved enquiry.

• Per accepted contract review.

• Per completed transaction.

• Per defect repaired.

• Per claim triaged.

A functional unit makes comparisons meaningful by defining the work delivered.

10. Software Carbon Intensity

Software Carbon Intensity (SCI) is a method for expressing the carbon impact of software relative to a defined functional unit. At a high level, it considers operational energy, the carbon intensity of the electricity and the embodied emissions allocated to the software system.

SCI for AI extends the approach to AI systems and their lifecycle. The quality of the result still depends on the quality and availability of the underlying data.

11. Embodied emissions

Embodied emissions are emissions associated with making and supplying physical equipment and infrastructure. For AI, that may include manufacturing servers, chips and other hardware. These emissions are different from the operational emissions associated with electricity consumed while the system is running.

Different methodologies may include or exclude embodied emissions, which can make apparently similar carbon figures difficult to compare.

12. Carbon intensity

Carbon intensity describes the greenhouse-gas emissions associated with producing a unit of electricity.

It can vary by country, region, season and time of day. The same amount of computing performed in two locations may therefore have different estimated carbon emissions. Lower operational carbon does not necessarily mean lower water use or lower total environmental impact.

13. Water withdrawal and water consumption

Water withdrawal is water taken from a source. Water consumption is the portion not returned to the original source in a readily reusable form.

Data centres may use water directly for cooling. Electricity generation may also have its own water footprint. Water figures can differ substantially depending on which boundary and definition are used.

14. Rebound effect

The rebound effect occurs when an improvement in efficiency lowers the effective cost of an activity and encourages more of that activity. AI inference may become cheaper per task while total AI usage rises faster.

Efficiency and absolute consumption must therefore be measured separately.

15. FinOps

FinOps is a management discipline for connecting technology usage with financial accountability. It developed around cloud expenditure.

For AI, it increasingly covers model costs, tokens, infrastructure, orchestration, agents and cost per business outcome. FinOps is not simply an exercise in cutting costs.

Its purpose is to make spending visible, attributable and connected to value.

📚 Further reading

• International Energy Agency: Key Questions on Energy and AI (2026)

  The IEA’s updated analysis of electricity demand from data centres and AI.

  It reports that global data-centre electricity demand increased by 17% in 2025 and estimates that consumption in AI-focused data centres grew substantially faster. Its central projection has total data-centre electricity consumption rising from approximately 485 TWh in 2025 to around 950 TWh in 2030.

  These are modelled estimates and projections, not direct measurements of every data centre. They should not be treated as guaranteed outcomes.

• International Energy Agency: Energy and AI (2025)

  The IEA’s broader assessment of electricity demand from data centres and AI, the sources expected to supply that electricity, grid constraints and the potential for AI to improve efficiency elsewhere in the energy system. Useful for understanding the physical infrastructure and power-system context beneath individual AI workloads.

• International Energy Agency: Energy and AI Observatory (2025–)

  The IEA’s data and analysis resource for tracking the changing relationship between AI, data centres and energy. Includes an interactive map of data centre hubs alongside associated power and digital infrastructure.

• Ofgem: Average gas and electricity use explained

  Ofgem’s explanation of typical domestic electricity and gas consumption in Great Britain.

• U.S. Energy Information Administration: Residential electric bills in Hawaii and Connecticut are highest in the United States (2025)

  The EIA reports that average monthly electricity consumption per US residential customer was 865 kWh in 2024, equivalent to approximately 10,380 kWh a year.

  The figure refers to residential electricity customers and varies materially by state, climate, home type and heating or cooling arrangements.

• German Federal Statistical Office: Electricity consumption of households by household size

  Destatis reports average German household electricity consumption of 3,383 kWh in 2021, including electricity used for heating, hot water, lighting and appliances. The source is official, but the data are older than the British and US benchmarks. The German-home equivalent should therefore remain rounded and dated rather than presented as a current exact average.

• Chao Yang, Mei Chen, Wanyang Zhou, Xinyue Zhang and others:

  The Role of Online Consumer Reviews in Shaping Chinese Consumers’ Purchase Intentions for Energy-Saving Household Appliances (2024)

  The peer-reviewed article states, citing Chinese official data, that per-capita residential electricity consumption increased from approximately 500 kWh in 2013 to nearly 1,000 kWh in 2023. It does not provide a direct national average per family household. The Chinese-home equivalent in the article therefore remains an illustrative calculation rather than an official household benchmark.

• National Bureau of Statistics of China: Main Data of the Seventh National Population Census (2021)

  The census reports an average family-household size of 2.62 people in 2020. Combining that census figure with approximately 1,000 kWh of residential electricity use per person gives an indicative figure of around 2,620 kWh per family household. The two inputs relate to different periods and statistical exercises. The resulting household comparison is therefore approximate and should not be described as an official current Chinese average.

• Agency for Natural Resources and Energy, Japan: 10 Questions for Understanding the Current Energy Situation (2024 edition)

  The Japanese government publication uses an “ordinary household” consuming 400 kWh per month, equivalent to 4,800 kWh a year, in its illustration of renewable-energy surcharges. It states that the model is based on household research conducted by the Ministry of Internal Affairs and Communications.

  This is a policy illustration rather than a definitive national household average. Therefore, the Japanese-home equivalent is remain explicitly illustrative.

• Green Software Foundation: SCI for AI - Software Carbon Intensity for Artificial Intelligence

  A consensus-based specification extending the Software Carbon Intensity methodology to AI systems. It covers stages including data preparation, training, deployment and inference, and distinguishes between provider and consumer measurement boundaries. The specification supports functional units suited to different AI systems. Its usefulness still depends on the availability and quality of underlying energy, hardware and emissions data.

• Green Software Foundation: SCI for AI Specification Ratified: Standard for Measuring AI Emissions Across the Lifecycle (2025)

  The formal announcement that the SCI for AI specification had been ratified. It describes the specification as extending the ISO-certified SCI methodology across the AI lifecycle. This is an announcement from the organisation responsible for the standard, not independent evidence that the methodology has already achieved universal industry adoption.

• Green Software Foundation: Tokens and Greens: Measuring the Impacts of Agentic AI (2026)

  A practitioner-oriented article on the energy, carbon and water implications of agentic AI. It argues for task-level measurement, model routing, caching, limits on retries and human checkpoints. This is guidance from an industry standards organisation, not a peer-reviewed empirical survey of corporate AI deployments.

• MLCommons: MLPerf Power

• A framework for measuring machine-learning performance and power consumption together. It supports comparison across data-centre, client, edge and embedded systems and emphasises measurement of full-system power rather than relying solely on nominal chip ratings. It is useful for infrastructure benchmarking and procurement. It does not establish whether a corporate AI workflow created useful or economically valuable work.

• MLCommons: MLCommons Power Working Group Presents MLPerf Power Benchmark at IEEE HPCA Symposium (2025)

  An explanation of the MLPerf Power framework, its measurement requirements and early benchmarking findings. MLCommons notes that power alone is not enough: performance, throughput, model quality and system-level consumption must be considered together.

• Longju Bai, Zhemin Huang, Xingyao Wang, Jiao Sun, Rada Mihalcea, Erik Brynjolfsson, Alex Pentland and Jiaxin Pei: How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks (2026)

  A preprint examining token consumption by eight frontier models performing agentic coding tasks. The authors report that the tested agentic tasks consumed around 1,000 times more tokens than their code-reasoning and code-chat comparisons. Runs on the same task varied by as much as 30 times, while greater token use did not consistently produce greater accuracy. The study concerns specific coding benchmarks, models and agent configurations. Its reported multipliers should not be applied indiscriminately to every enterprise agent or business workflow.

• Lola Solovyeva and Fernando Castor: Towards Green AI: Decoding the Energy of LLM Inference in Software Development (2026)

  A preprint separating the energy used during the input-processing, or prefill, stage from the output-generation, or decoding, stage of language-model inference. In the models and coding benchmarks tested, the authors found that excessive output could materially increase energy consumption. Their suppression technique produced reported savings of between 44% and 89% without reducing measured code-generation accuracy. Those percentages relate to the tested open-weight models and controlled software-development tasks. They should not be generalised to proprietary models or unrelated corporate workflows.

• Warren Johnson: The Compression Paradox in LLM Inference: Provider-Dependent Energy Effects of Prompt Compression (2026)

  A preprint examining prompt compression across three API providers, five benchmarks and several compression ratios. The study found that reducing input length did not reliably produce a better energy-quality trade-off. In some tested cases, compression reduced answer quality or caused output expansion. The paper estimates the energy use of proprietary services using a token-based proxy calibrated against local measurements. It does not directly observe the providers’ physical infrastructure.

• FinOps Foundation: FinOps for AI: Tools & Services Considerations (2026)

  Practical guidance on understanding AI costs across training, tuning, inference, orchestration, observability and operations. It argues that the real unit of measurement is the economics of the use case, including whether a tool is necessary, correctly scoped and worth its total cost. This is practitioner guidance rather than an accounting or regulatory standard.

• Linux Foundation: Linux Foundation Announces the Intent to Launch the Tokenomics Foundation to Establish Open Standards for AI Cost Management (2026)

  The June 2026 announcement sets out plans for a vendor-neutral body focused on AI infrastructure economics, token efficiency, benchmarks and cost-management standards. The initiative is evidence of demand for common financial measurement as generative and agentic AI move into production. At the time of writing, it is an announced foundation and work programme, not a completed or universally adopted standard.

• European Commission: Targeted Consultation on Measuring Energy Consumption and Emissions of AI Models and Systems (2026)

  The consultation seeks evidence to support a practical measurement framework for the energy-related objectives of the EU AI Act and the possible design of an AI energy and emissions label. The Commission notes that providers of general-purpose AI models must document known or estimated energy consumption under Annex XI of the AI Act. The consultation also demonstrates that the measurement framework remains under development. It does not establish that a final, comprehensive and universally accepted methodology already exists.

• National Institute of Standards and Technology: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (2024)

  NIST’s cross-sectoral companion to the AI Risk Management Framework for generative AI. It identifies risks and suggested actions across governance, mapping, measurement and management. It is relevant to the article’s discussion of testing, monitoring, third-party dependencies, human oversight, incident response and documentation. The framework is voluntary guidance. It is not an insurance policy, underwriting standard or guarantee that an AI system will be insurable.

• The Geneva Association: Gen AI Risks for Businesses: Exploring the Role for Insurance (2025)

  An insurance-industry research report examining generative-AI risks across operational, cybersecurity and privacy, ethical, regulatory, reputational, workforce and environmental domains. It considers corporate awareness of those exposures and the potential role of insurance. The report is useful to understand why AI losses may cross several established insurance classes. It does not establish that every identified AI risk is currently transferable or commercially insurable.

• Jones Day: “A-Eye” on Coverage: Maximizing Insurance for AI Risks Amid Emerging Exclusions (2026)

  Policyholder-side legal commentary on emerging AI exclusions and the possible application of cyber, errors and omissions, directors’ and officers’, crime, media and other policies. It advises insureds to examine renewal terms carefully and negotiate exclusions where possible. This is legal commentary rather than a market-wide empirical survey. It supports the statement that AI exclusions are emerging, but it does not prove that most carriers or policies exclude all AI-related losses.

• Alex Leung, Rex Zhang, Ervin Ling, Kentaroh Toyoda and SiewMei Loh: The Insurability Frontier of AI Risk: Mapping Threats to Affirmative Coverage, Silent Exposures, and Exclusions (2026)

  A preprint mapping 55 AI threat classes against 26 publicly described insurance products, endorsements and exclusion regimes. The paper identifies affirmative coverage, silent exposure, actively excluded risks and risks that may sit outside conventional private insurance. Its analysis is based on public carrier materials rather than complete executed policy wordings or adjudicated claims. Its conclusions should therefore be treated as a structured market map, not evidence that any particular claim would be paid.

• Quanyan Zhu: Insurance of Agentic AI (2026)

  A preprint examining how autonomous planning, tool use and execution create exposures across cyber, professional liability, product liability and other established insurance classes. It proposes that agentic-AI risk may ultimately require coordinated layers of coverage, explicit allocation mechanisms, dedicated aggregates, improved telemetry and clearer governance rather than one universal AI policy. This is a conceptual academic framework published as a preprint.

• Dr Henri Winand: The Hidden Appetite of AI (2025)

  An earlier Something for the Weekend article examining the physical energy and water appetite behind AI systems. The article introduced the argument that apparently weightless AI services rest on power-intensive physical infrastructure.

• Dr Henri Winand: The Digital Employee Was Never $200 a Month (2026)

  The argument that a subscription is only the entry price and that the economic cost of machine work includes tokens, inference, infrastructure, integration, verification and risk.

• Dr Henri Winand: Rent the Intelligence. Own the Proof. (2026)

  An examination of cloud versus local AI, larger versus smaller models and the need for evidence records in serious corporate workflows.

• Dr Henri Winand: When Tokens Become Kilowatts (2026)

  Last week’s article on AI tokenomics, machine-work consumption and the physical infrastructure beneath the token.

Last week on Something for the Weekend

Last week’s piece looked at the new tokenomics:

• Crypto used tokens to make scarcity, incentives and governance visible.

• AI uses tokens to make consumption visible.

The article followed the AI token down through inference, compute and data-centre infrastructure to the electricity system underneath. It argued that the company which cannot see token consumption cannot see the cost of machine work.

However, tokens are only the basic meter.

The better meter is useful, verified work.

This week asks what happens after companies install that meter across thousands of digital co-workers. The next advantage will not come from producing the most machine activity, it will come from producing the most verified value with the least avoidable waste.

A CFO receives three updates before breakfast. The first says the company’s AI tools are lifting productivity. The second says the AI budget is running ahead of plan. The third says a model gave a confident answer in a legal workflow, but missed the point because the relevant context never made it into the answer.

The CIO explains that the issue is tokens.

The CFO thinks: not crypto again.

But this time the token is not a coin, a speculative asset, a governance right, or a claim on a protocol. It is a unit of text and data processed by an AI model. A short prompt consumes a few. A long document review consumes more. A coding agent that reads files, writes code, tests, fails, retries and calls tools can consume a great many.

The company thought it had bought software but it had bought access to a meter.

That is why the word “tokenomics” is becoming useful again, but not in the way most people remember it. The crypto version was about issuance, scarcity, incentives, speculation and governance. How many tokens exist? Who receives them? Are they minted, burned, staked, locked, vested or diluted? Does the token buy access, status, votes, yield, collateral or hope?

AI tokenomics is different. It is about consumption. Who uses the tokens? Who pays for them? Who controls the meter? Which workflow burns through them? Which agent can keep running in the background? Which model is too large for the job? Which answer needs verification? Which user sees the value but not the bill?

At first, the two token stories look unrelated:

One belongs to crypto.

The other belongs to AI.

But the underlying story is the same: both sit on top of machines that convert electricity into economically useful digital output. In proof-of-work crypto, electricity is converted into cryptographic work, network security and block rewards. In AI, electricity is converted into inference, output, tool use, agent work and, sometimes, confident error.

The token is the surface unit and the substrate is compute. But, the bottleneck is increasingly power. That is the new tokenomics.

Not coins.

Kilowatts.

The Boardroom point

The next tokenomics debate will not be about whether a coin goes up. It will be about whether machine work creates value before it consumes too much cost, power, attention and trust.

Enterprises are starting to discover that AI is not a normal software budget. It behaves more like cloud consumption, professional services, outsourcing, infrastructure and labour all at once. There is a subscription. Then there is usage. There is a model. Then there is context. There is a prompt. Then there are retries. There is output. Then there is verification. The visible price is the entry ticket and the economic price is the work system behind it.

This is a direct extension of a point I made recently in Something for the Weekend: the digital employee was never $200 a month. The subscription was the entry price. The full cost included tokens, inference, context, API calls, cloud infrastructure, integration, verification and risk. Then the next problem arrived. Where does the intelligence run? In the cloud, locally, or across a hybrid stack? What does it cost? Can the company prove why the answer was given?

This week’s point sits underneath both:

The token is the meter.

The company that cannot see token consumption cannot see machine-work cost. The company that cannot see machine-work cost cannot measure AI productivity properly. The company that cannot measure AI productivity properly is likely to confuse activity with value.

Prompts sent are not value. Documents generated are not value. Agents launched are not value. Tokens consumed are not value. The useful measure is verified machine work.

Did the AI do something useful? Was it correct? Was it used? Did it reduce old work? Did it increase revenue? Did it lower risk? Did it shorten cycle time after review? Did it reduce rework? Did it create a decision that can be defended later?

The token is not the answer.

It is the meter attached to the question.

The old tokenomics

Crypto tokenomics began with a powerful idea. Digital systems could create economic rules. A token could represent access, payment, governance, collateral, participation, status, scarcity, or a claim on a future system. Code could set issuance schedules. Protocols could define incentives. Communities could coordinate without a traditional company at the centre.

Some of that was genuinely interesting, but once hype set in, much of it was not.

Too often, the token came first and the economy came later:

• A white paper promised utility.

• A supply schedule created scarcity.

• A community created momentum.

• A market price created excitement.

• The token moved before the system had useful work to do.

That was the failure mode with scarcity confused with value.

A limited supply of something of no value does not make it useful. A governance token without meaningful governance is not participation. A token that exists mainly to sell the next token is not an economy, it is a financing mechanism with a narrative attached.

This does not mean tokenomics was meaningless. It means the serious version was never just about coins. It was about incentive design: who contributes, who verifies, who governs, who gets rewarded, who takes risk, who can exit, who gets diluted, who can change the rules, and who benefits when the system scales.

Those are good questions.

They are now returning through a different door:

Not through crypto speculation.

Through AI consumption.

The new tokenomics

AI reverses the direction of the token. The AI token is not usually something the company owns. It is something the company spends.

Every prompt spends tokens. A simple question spends a few. A long contract spends more. A large context window spends more again. An agent that reads files, searches systems, calls tools, writes code, tests it, fails, retries, produces a report and then revises it can spend far more than the user realises.

This changes the economics of work. A human employee has limits. They can only work so many hours. They appear slow compared with software, but they are bounded:

They can be questioned.

They can explain what they did.

They can be trained, managed and held responsible.

A digital worker is different. It can run continuously. It can generate more output than the organisation can absorb. It can create plausible work that still needs expert review. It can multiply mistakes. It can shift costs from headcount to compute. It can appear cheap until usage scales. But, it also has no real accountability.

That is why AI tokenomics is not just a technical topic. In reality, it is a management topic:

The question is no longer simply how much the model costs per token, the better question to ask is how many tokens it takes to produce a verified outcome.

That distinction is important. A cheap model that produces uncertain work may become expensive if senior people have to spend hours checking it. A more expensive model may be cheaper if it reaches a reliable answer faster. A small model may beat a large model if the task is narrow, confidential and repetitive. A local system may beat a cloud system if the data should not move. A frontier model may still be worth the cost if the problem needs broad reasoning.

The cheapest token is not always the cheapest work:

The energy underneath

The AI token appears weightless. However, it is not.

Beneath the token lies inference. Inference, in turn, relies on compute. Compute depends on GPUs, memory, networks, cooling, land, water, grid connections, power contracts and data-centre finance. The browser makes AI feel like software. The infrastructure makes it look more like heavy industry.

That is why the data centre is becoming a digital factory.

It takes power, machines, cooling, capital and skilled operations, then turns them into digital output: tokens, answers, simulations, classifications, code, proofs, decisions and agent work. The product is weightless. The factory is not.

This is where the crypto comparison becomes useful, but only if handled carefully.

In proof-of-work crypto, the energy link is explicit. Electricity is converted into cryptographic work, network security and block rewards. Bitcoin made that visible because mining was part of the system’s public logic.

AI is different. No one is mining AI tokens into existence. AI tokens are not coins. They are not scarce digital assets. They are metering units for model work. But the physical link still exists. Every token processed by a model requires compute. Compute requires energy. Energy requires infrastructure.

So the better comparison is this:

Crypto made digital scarcity visible.

AI makes digital consumption visible.

Both return us to the same hard reality: the digital economy is not detached from the physical world. The front end is a token, the back end is machine time and the limiting input is increasingly power.

This is why “cost per token” and “tokens per watt” are becoming strategically important. Those phrases sound technical, but they are commercial. If more billable AI work fits inside the same power envelope, margins improve. If a company burns tokens without producing verified value, the cost base expands without the productivity gain.

AI will not only be judged by model capability. It will be judged by energy-to-output efficiency:

What if the wrong compute gets built?

The easy version of the AI bubble story is that too many data centres get built, demand disappoints, and the whole thing becomes another infrastructure mistake.

That’s the crude story some are happy to write.

The more interesting possibility is that the world builds too much of one kind of compute just as AI work starts to split into different places. Some work will still need frontier models in the cloud. Some work will need fast inference close to the user. Some work will need smaller models running locally because the data are sensitive, the task is narrow, the latency matters, or the company does not want every question to become an external cloud call.

This does not mean cloud AI goes away. It means cloud AI becomes part of a portfolio.

A training cluster, an inference cloud, a local model, and an edge device are distinct concepts. A general-purpose frontier model differs from a smaller model embedded within a legal, insurance, banking, manufacturing, or customer-service workflow.

So the real overbuild risk may not be that nobody needs AI compute. The risk may be that the wrong compute is built in the wrong place, under the wrong pricing assumptions, for workloads that later become cheaper, smaller, more local or more specialised.

That is a much more useful boardroom question.

If small models and local inference become good enough for many enterprise tasks, the largest models do not disappear, they become the expensive specialists. They handle the hard work: broad reasoning, difficult coding, complex research, high-end synthesis and tasks where the best available model is worth the cost.

Routine work may route elsewhere:

• Smaller models.

• Private deployments.

• Local machines.

• Edge devices.

• Retrieval-first systems.

• Domain-specific agents.

• Workflow-specific tools.

That would change token economics. Cloud tokens would not vanish, but they could become cheaper, more competitive and more carefully rationed. Cost per token may fall because of better chips, better software, better utilisation, model optimisation and competition. At the same time, total AI usage could still rise because lower costs make more machine work economically viable.

That is the awkward (Jevons) paradox: cheaper intelligence may not reduce the AI bill, it may increase the amount of AI work. For serious users, that is not necessarily bad. A fall in the cost of machine work could make AI more useful across more business processes.

The problem is not cheaper tokens. The problem is using them to generate more unverified activity.

This is where boards should be precise. The question is not whether the company should use cloud AI, local AI or smaller models. The answer will usually be all three.

The question is which work belongs where.

A broad strategic question may need a frontier model. A confidential contract review may need local control. A repetitive compliance check may need a smaller model. A high-volume customer workflow may need an inference cloud. A consequential decision may need human approval and an audit receipt.

But the deeper point remains: once the world builds more power-dense digital factories, the question becomes what those factories can run next?

Inference.

Simulation.

Scientific computing.

Private AI.

Enterprise agents.

Cryptographic proof systems.

Tokenised contracts.

Distributed ledgers.

Future machine-to-machine commerce.

Maybe even some crypto mining where the economics and equipment make sense.

That caveat is important. A GPU-heavy AI facility is not automatically a bitcoin mine. Bitcoin mining is now heavily ASIC-driven, while AI data centres are built around GPUs, accelerators, networking, storage, cooling and service levels suited to different workloads. Some conversions will be impractical. Some will be uneconomic. Some sites will simply be wrong.

Still, if there is a bubble, it may leave behind something useful: a larger base of digital factories.

The future is unlikely to be less AI: it is more likely to be more AI, running in more places, with more discipline about what each token is for.

Some capital may be misallocated. Some vendors may disappoint. Some sites may age badly or be poorly located. But the broader machine-work layer is still likely to become cheaper, more distributed and more useful.

The winners will not be the companies that simply buy the most AI.

They will be the companies that route the right work to the right machine, at the right cost, with the right proof.

The agent clock card

This is where crypto and distributed ledgers return in a more useful form, not as speculation, as accounting infrastructure.

A human factory worker used to clock in with a time card. A consultant has a timesheet. A trader had an order blotter. A lawyer has a file note. A bank has transaction logs. A board has minutes.

An AI agent needs its own version of that, because the agent can act.

An enterprise AI agent may read documents, access systems, call tools, write code, draft contracts, respond to customers, generate reports, approve workflow steps, raise tickets, move data and trigger downstream actions. If it can do those things, the company needs to know what it did, when, under whose authority, against which budget, with which data, using which model, and with what result.

That is where distributed ledgers become interesting again. Not every AI audit trail needs a public blockchain. In many enterprise settings, the more plausible answer will be permissioned infrastructure, private ledgers, cryptographic receipts, hash commitments, tamper-evident logs or selectively disclosed proofs.

Most Fortune 500 companies will not want sensitive workflow data sprayed onto a public chain. The point is not public visibility. The point is controlled auditability.

The concept is simple: give every agent a clock card.

The clock card should show the agent identity, delegated authority, budget, model used, prompt class, data accessed, tools called, tokens consumed, outputs produced, abstentions, exceptions, human approvals and final decision receipt.

Importantly, some of that information may stay private and some may be hashed. Yet, some may be visible to auditors, to regulators, insurers or counterparties under defined conditions.

The point is not to turn every AI action into a tradable token, it is to make machine work auditable.

That is the useful convergence of AI and distributed ledgers:

AI creates work at machine speed.

Distributed ledgers can help create a tamper-evident record of that work.

Smart contracts may then enforce budgets, approvals, access rights and stop conditions.

In short:

An AI agent without a meter is a risk.

An AI agent with a meter, a budget and a receipt starts to look like manageable machine labour.

That is a boardroom-grade idea.

Engagement and dependency

There is another uncomfortable layer. AI systems do not exist outside commercial incentives.

If a vendor only sells raw tokens, the business risks moving toward commodity economics: electricity, data centres, chips, model training, inference, cooling, depreciation and utilisation. That can be a very large business. It is not necessarily the best-margin business.

The better business is to keep the user inside higher-value workflows: coding environments, legal tools, research systems, insurance applications, banking workflows, analytics layers, knowledge platforms, enterprise agents, procurement systems and customer service infrastructure.

That is where the supermarket analogy becomes useful. The raw model is not the whole business. It is the shelf space. The real margin sits in what gets placed on the shelves, how often customers return, how much trust they place in the store, and which value-added services are wrapped around the visit.

“Stack them high and sell them cheap” can be a good business. The better business is when the platform becomes the place where the customer does more of the work.

That raises a serious issue for business when it comes to productivity and cost:

A model that gives a confident answer may be more engaging.

A model that keeps the user asking follow-up questions may consume more tokens.

A model embedded into a workflow may create dependency.

A model that is helpful enough to keep the user inside the system may become the new operating layer.

None of that proves bad faith. It does not mean AI vendors want hallucinations. It does mean product design, benchmark design and commercial design matter.

If the system rewards answering, the model will answer. If the benchmark penalises uncertainty, the model will guess. If the user wants fluency, the product will provide fluency. If the vendor monetises usage, continued usage matters.

That is not a conspiracy. It is economics. When billions are being invested in models, infrastructure and distribution, usage matters.

The safer model may be the one that knows when to say: I do not know.

Guessing burns tokens too

This is where hallucination becomes part of tokenomics. A wrong answer is not just a quality problem, it is a consumption problem.

A confident wrong answer can trigger follow-up prompts, document checking, expert review, correction cycles, duplicated work, legal uncertainty, customer confusion and reputational risk. The hallucination consumes tokens once when it is generated, then it consumes more tokens, time and attention when the organisation has to work out whether it was wrong.

That is verification debt with a meter attached.

Recent research on hallucination makes the incentive problem sharper. The central point is that many accuracy-based benchmarks reward guessing more than abstention. A model that says “I do not know” receives no credit. A model that guesses has at least some chance of being right. So the training and evaluation environment can structurally under-reward uncertainty.

That matters in low-risk contexts but it truly matters in serious work:

• A chatbot making up an acronym is irritating.

• A legal tool making up a limitation point is dangerous.

• A banking system misreading a policy is dangerous.

• An insurance workflow missing an exclusion is dangerous.

• A board paper using the wrong assumption is dangerous.

So, a model that guesses confidently may keep the interaction alive but it can create costly false confidence.

The enterprise answer is not to ban AI.

The answer is to design for abstention, verification and routing.

Some questions should go to a frontier model. Some should go to a smaller model. Some should go to retrieval first. Some should go to a human. Some should stop.

Yet again, “I do not know” may be one of the most valuable outputs in enterprise AI because it stops the meter before false confidence becomes expensive.

The CFO version

For the CFO, AI tokenomics is a cost-structure problem.

Traditional software was easier to budget: seats, licences, renewals, implementation, support and annual uplift. AI changes the shape. There may still be seats and subscriptions, but the real economic motion comes from use. The company pays for access and then pays again through consumption, directly or indirectly.

The difficult part is that the people creating the cost may not see the bill. A developer running a coding agent may not see the cloud cost. A lawyer asking for long document reviews may not see the token cost. A business unit launching an AI pilot may not see the integration cost. A customer-service team automating replies may not see the verification cost. A board approving “AI productivity” may not see the full cost of machine work.

That is why the finance team needs a new language:

Cost per token is useful but incomplete. Cost per verified outcome is better.

Cost per avoided hour may be useful, but only if the old work actually disappears.

Cost per decision may be useful, but only if the decision can be reconstructed.

Cost per agent run may be useful, but only if the agent achieves the task.

The CFO should ask where tokens are being consumed, which users or agents can spend them without approval, which costs are fixed or variable, which vendors can change usage limits or pricing, and which AI costs are hidden inside software, cloud, consulting, legal, compliance or security budgets.

But the most important question is simpler:

The company has not proved AI productivity because usage is rising. It has proved AI productivity when useful output rises after the full cost of tokens, compute, verification, integration, rework, risk and accountability is counted.

The CIO version

For the CIO, AI tokenomics is an architecture problem. That is because the company needs routing rules:

• Which work should use a frontier model?

• Which work should use a smaller model?

• Which work should run locally?

• Which work can go to the cloud?

• Which work needs retrieval first?

• Which work needs a retained evidence record?

• Which work needs human approval?

• Which work should not be automated?

That last question is not anti-technology, it is simply architecture discipline.

The lazy version of AI is to send everything to the largest available model and hope scale solves the problem. That is expensive and weak governance.

A large model may be right for broad reasoning, coding, research and synthesis. It may be overkill for document classification, invoice matching, policy checks, contract comparison, claims routing, meeting summaries or structured extraction. A small model doing a narrow job well may be better than a large model doing the same job expensively and inconsistently.

This is where tokenomics becomes operational design: the goal is not to minimise tokens in isolation, it is to minimise waste:

Use enough model for the task.

Use enough context for the answer.

Use enough verification for the risk.

Use enough evidence for the decision.

Use enough human judgement for accountability.

The CIO who treats tokens as a technical billing unit will miss the point, tokens are becoming one of the ways machine work becomes visible.

The Board version

For the board, tokenomics is an accountability problem. Boards do not need to understand every model parameter, context window or inference stack. They do need to understand whether the company is building a machine-work cost base it can control.

A useful board discussion should cover ten points:

1. Where is AI being used in production?

2. Which AI systems are metered by tokens, API calls, compute, credits or usage tiers?

3. Which workflows consume the most machine work?

4. Which agents can act repeatedly without human intervention?

5. Which AI outputs require human approval?

6. Which AI systems can say “I do not know”?

7. Which AI systems are being rewarded for guessing?

8. Which AI decisions can be reconstructed later?

9. Which vendors control pricing, rate limits and model access?

10. Which AI gains have reached the P&L after the full bill is counted?

The board should also ask a harder question:

That question will become more important as AI moves from chatbots into workflow systems. The risk is not simply that the model is wrong. The risk is that the company builds daily operations around systems whose incentives it has not inspected.

An AI answer without a record is hard to challenge. An AI workflow without cost visibility is hard to manage. An AI system without abstention is hard to trust.

An AI strategy without token discipline is just cloud spending with better language.

The insurance question

There is another reason why understanding the meter is important.

Boards will want AI risk to be insurable. Insurers will want to know whether that risk is covered, excluded, sublimited or separated into a different coverage line.

That is not surprising. Once AI systems produce operational, legal, cyber, financial and reputational losses, insurers will ask what exactly is being covered. Is it cyber? Is it technology errors and omissions? Is it professional liability? Is it crime? Is it operational risk? Is it a new AI-specific exposure?

One emerging concern is LLMjacking, where attackers hijack AI accounts or credentials and burn through compute at someone else’s expense. That is a natural consequence of metered machine work.

If access to AI systems becomes valuable, attackers will steal it. If tokens represent spend, attackers will burn them. If agents can act, attackers will redirect them. If models sit inside business workflows, failures will create claims.

This creates a new Board and insurance problem. The loss is not always the wrong answer. Sometimes the loss is the machine doing too much of the wrong work.

A stolen password used to mean unauthorised access. A stolen AI credential may mean unauthorised consumption, unauthorised output, unauthorised decisions, unauthorised data movement and unauthorised reliance.

That is a different risk profile.

Policy language will have to catch up. Sublimits may be presented as clarity. They may also become the place where clients discover that the AI loss they assumed was covered is only covered up to a narrow ceiling.

The practical point is simple:

If AI work is metered, AI loss can be metered too and so, who pays when the meter is hijacked?

The weekend takeaway

The word token has done a strange journey:

In crypto, the token was sold as a digital asset.

In AI, the token has become a unit of machine work.

In both cases, the token makes part of the machine economy visible. Crypto showed how digital systems could create scarcity, incentives, speculation and governance. AI shows how digital systems create consumption, cost, verification debt and dependency.

The shared lesson is not that crypto tokens and AI tokens are the same, it is that tokens reveal the economics of the system underneath:

At the bottom of that system sits power.

If the AI build-out overheats, some capital will be lost. That is how investment cycles work. But the more interesting question is what gets left behind: power-dense sites, grid connections, cooling systems, fibre, operating teams, compute markets and a broader industrial base for machine work.

AI may offer the opportunity. Overbuild, if it happens, could make compute cheaper, inference more available, agent work more measurable, and digital infrastructure more reusable. That is the positive case.

The next step is discipline: the future digital worker needs a meter, a budget and a receipt:

The token is not the product.

The token is not the intelligence.

The token is not the proof.

The token is the basic meter.

The better meter is useful, verified work.

The company that truly understands the meter will manage AI as work.

The company that does not will manage AI as magic until the invoice arrives.

Leave a comment

🧠 Techie Corner for the Non-Techies

1. Token

A token in AI is a unit of text or data processed by a model. Short prompts use fewer tokens. Long documents, long conversations, codebases and agentic workflows use more. Tokens matter because many AI systems are priced, limited or managed by how much input and output the model processes.

This is different from a crypto token. A crypto token is usually a digital asset or claim inside a blockchain-based system. An AI token is usually a metering unit for machine work:

Same word, different economics.

2. Tokenomics

Tokenomics means the economics of tokens. In crypto, tokenomics usually means supply, issuance, utility, incentives, staking, vesting, burning, governance and market behaviour. In AI, tokenomics should mean something different: how tokens are consumed, priced, routed, verified and converted into useful work.

Crypto tokenomics asks how a digital system creates and allocates value.

AI tokenomics asks how a digital system consumes cost to produce output.

3. Proof-of-work

Proof-of-work is the system used by Bitcoin and some other blockchains to secure the network. Miners use specialised hardware to perform computational work. The successful miner earns the right to add a block and receives rewards according to the protocol rules.

The important point for this article is that proof-of-work makes the energy link visible. Electricity is converted into cryptographic security and block rewards.

That does not apply equally to all crypto. Proof-of-stake systems work differently and are much less energy-intensive.

4. Inference

Inference is what happens when an AI model runs to answer a prompt or complete a task. Training is the expensive process of building or improving the model. Inference is the repeated cost of using it.

Most companies will not train frontier models. Many will pay for inference every day.

That is where the meter appears.

5. Context window

The context window is the amount of information an AI model can consider at one time. A larger context window lets the model read more documents, code or conversation history. It can also cost more because the model must process more material.

That is why “give the AI everything” is often inefficient. Better systems retrieve the relevant material first, then ask the model to work from that smaller set.

6. Agent

An AI agent is a system that can take multiple steps toward a goal. It may search files, call tools, write code, test outputs, revise its answer and try again.

Agents can be useful because they do more than answer one prompt. They can also become expensive because each step consumes tokens and compute.

The key question is not whether the agent ran.

The key question is whether it completed useful work within a controlled cost and risk boundary.

7. Distributed ledger

A distributed ledger is a shared record of transactions or events maintained across multiple systems or parties. It can be designed to make records tamper-evident and independently verifiable.

For AI agents, the useful idea is not speculative trading. It is auditability. A ledger can help record that a machine action happened, under what authority, and with what evidence, without necessarily exposing the underlying confidential material to everyone.

8. Agent clock card

An agent clock card is a simple way to describe an audit record for AI work.

It should record the agent identity, authority, budget, model used, prompt class, data accessed, tools called, tokens consumed, output produced, abstentions, exceptions, human approvals and final decision receipt.

The point is not to humanise the agent, it is to control the work.

9. Verification debt

Verification debt is the checking burden created by machine-generated output. AI can produce more drafts, summaries, code, analysis and recommendations than an organisation can easily review.

If the organisation generates faster than it can verify, the apparent productivity gain becomes a bottleneck.

Verification debt is where cheap generation becomes expensive.

10. Tokens per watt

Tokens per watt measures how much AI work can be produced from a unit of power. AI is not weightless. Behind the model sits hardware, power, cooling, memory, networking and data-centre infrastructure.

As AI scales, energy efficiency becomes part of the economics of intelligence.

11. LLMjacking

LLMjacking means attackers hijack AI accounts, credentials or access and use someone else’s AI resources.

The result may be unauthorised token consumption, unauthorised compute spend, data exposure, operational disruption or downstream liability.

As AI becomes metered work, stolen access becomes a financial and operational risk.

📚 Further reading

• International Energy Agency: Energy and AI (2025)

  Estimates global data-centre electricity consumption at about 415 TWh in 2024 and projects around 945 TWh by 2030 in its base case.

• Reuters: Australia’s Megaport secures four new AI infrastructure contracts, to raise $594 million (2026)

  Reports Megaport’s new AI infrastructure contracts and capital raise, supporting the argument that AI demand is not only about giant training clusters but also about globally distributed, latency-sensitive inference capacity.

• Core Scientific: Core Scientific to Provide Approximately 200 MW of Infrastructure to Host CoreWeave’s High-Performance Computing Services, Capturing Significant AI Compute Opportunity (2024)

  Power-footprint convergence between crypto-mining infrastructure and AI/HPC hosting: Core Scientific announced that it would provide approximately 200 MW of infrastructure to host CoreWeave’s high-performance computing operations.

• CoreWeave: CoreWeave to Acquire Core Scientific (2025)

  CoreWeave proposed a $9 billion acquisition to secure Core Scientific’s 1.3 GW data center footprint and power capacity for future AI infrastructure. Though the all-stock deal was terminated by Core Scientific in late 2025 due to a lack of stockholder approval, it proved that direct power and data center control have become highly prized strategic assets.

• NVIDIA: Rethinking AI TCO: Why Cost per Token Is the Only Metric That Matters (2026)

  Cost-per-token framing: NVIDIA defines cost per token as the all-in cost to produce each delivered token and argues that it is a better AI infrastructure metric than raw compute cost or FLOPS per dollar.

• NVIDIA Developer Blog: Scaling Token Factory Revenue and AI Efficiency by Maximizing Performance per Watt (2026)

  NVIDIA links AI data-centre economics to power, land, shell capacity, token throughput, performance per watt and revenue per megawatt. Again, treat this as vendor-positioning rather than neutral analysis.

• Nature: Evaluating large language models for accuracy incentivizes hallucinations (2026)

  The paper argues that next-word prediction and accuracy-based evaluations can reward unwarranted guessing, and that dominant headline metrics may reward guessing over admitting uncertainty i.e. “I do not know” can be economically and operationally valuable.

• OpenAI: Why Language Models Hallucinate (2025)

  Argues that language models hallucinate partly because standard training and evaluation procedures reward guessing over acknowledging uncertainty.

• Aegon: Auditable AI Content Access with Ledger-Bound Tokens and Hardware-Attested Mobile Receipts (2026)

  The paper proposes ledger-bound tokens, a Merkle-tree append-only transaction ledger, signed provenance event logs and hardware-attested compliance receipts for AI content access. It is early-stage research, not a market standard.

• From Agent Traces to Trust: Evidence Tracing and Execution Provenance in LLM Agents (2026)

  The paper argues that final-answer accuracy is not enough for LLM agents and that process-level accountability requires evidence tracing and execution provenance.

• NIST: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (2024)

  Useful governance reference. NIST describes this as a cross-sectoral companion resource to the AI Risk Management Framework for generative AI, aimed at helping organisations incorporate trustworthiness considerations into AI design, development, use and evaluation.

• Dr Henri Winand: The Digital Employee Was Never $200 a Month (2026)

  Why AI labour should be treated as metered machine work rather than fixed-price employment. The piece defines the digital-employee cost base as including tokens, inference, context, API calls, cloud infrastructure, integration, verification and risk.

• Dr Henri Winand: Rent the intelligence. Own the proof. (2026)

  Article on cloud AI, local AI, smaller models, compute cost and audit trails. It supplies the continuity for the argument that serious AI decisions need control, cost visibility and a decision receipt.

• Dr Henri Winand: The Future Is Being Built Without a Front Door (2026)

  Article on AI infrastructure, participation, work, family formation and the need for people to see a route into the future being built around them.

Last Week on Something for the Weekend

Last week’s piece looked at the participation bargain underneath the AI build-out.

The argument was that a country can build data centres, fund AI models, celebrate productivity gains and still lose the confidence of the people expected to live inside that future. But growth needs a front door.

Young people need a route into work. Families need confidence. Communities hosting infrastructure need visible benefit. Employees need AI to improve work rather than simply increase pressure. Boards need productivity gains that reach the P&L without destroying the learning systems that create future talent.

The useful question was:

Can people still see a route into the future being built around them?

That piece looked at the human and social bargain.

This week looks at the machine-work meter underneath it: the token.

The future may be built with AI, but the invoice will arrive in tokens.

A woman in Nottingham tells the BBC that the world does not feel like a nice place to bring children into. A trainee nurse in London says she would love a larger family, but cannot afford to give another child what her existing children already need. A 23-year-old graduate applies for more than 400 jobs and gets one interview for a cleaning role. A government-commissioned review warns that Britain is at risk of a “lost generation”.

The same pressure appears elsewhere.

In the United States, provisional CDC data show about 3.6 million births in 2025, down 1% from 2024, with the general fertility rate at 53.1 births per 1,000 women aged 15-44. In France, INSEE reported 645,000 births in 2025, 24% fewer than in 2010, and more deaths than births for the first time since the end of the Second World War. In Germany, Destatis reported a fertility rate of 1.35 children per woman in 2024.

In Japan, the number of children has fallen for decades. Japan’s child population is now at record lows. In South Korea, the OECD has described the country’s fertility rate as the lowest in the world, around 0.7 in 2023. In India, the problem looks different: a large young population, but an India Employment Report warning that educated young people face high unemployment and that many young people, particularly young women, are outside education, employment or training. Across the EU, Eurostat reported that 11% of 15-29-year-olds were neither in employment nor in education or training in 2025, and that youth unemployment was 15.4% in March 2026.

At the same time, the AI infrastructure build-out continues. Data centres, power contracts, GPUs, fibre, cooling, private capital, grid connections, planning applications and political fights are becoming part of the ordinary economics of the next decade.

That is the part Boards need to understand.

A country can build data centres, fund AI models, celebrate productivity gains and still lose the confidence of the people expected to live inside that future. That loss of confidence does not arrive as one event. It appears in private decisions: having fewer children, delaying a family, leaving education with no clear route into work, taking a job that does not build a career, using AI because it is necessary while trusting it less, and resisting infrastructure that claims to support national progress.

The good news is that this can be fixed.

The route back starts with a simple recognition: growth needs a front door. People need to see how they enter, contribute, learn, earn, build families, own assets and benefit from the technology being deployed around them.

The Boardroom point

Growth depends on participation. Technology can raise output, reduce costs, make small teams more productive and create demand for new power systems, grid upgrades, chips, cooling, batteries, modular construction and other hard technologies. Data centres may even pull technologies down their cost curves by giving them an early buyer with deep demand. Packy McCormick made that argument recently in Not Boring. It is a serious argument and should not be dismissed.

The hard part is the social bargain. Growth becomes fragile when people cannot see their route into it. In the UK, births are at their lowest level in nearly 50 years and more than 1 million young people are Not in Education, Employment or Training (the so called NEETs). In the US, births remain near historic lows. In France, the old demographic confidence has weakened. In Germany, Japan and South Korea, ageing is already a national economic constraint. In India, the challenge is whether education turns into good work at sufficient scale. Across the EU, the NEET and youth unemployment numbers remain material.

The pattern is not identical across countries and should not be forced into one cause. France’s fertility decline is not the same as India’s graduate underemployment. Japan’s ageing problem is not the same as Britain’s NEET problem. America’s childcare and healthcare costs are not Germany’s industrial workforce problem.

But the Boardroom relevance is common. The route from childhood to education, work, household formation, family formation and long-term consumption is becoming less reliable in many large economies. That route sits beneath demand, labour supply, taxation, political stability and corporate growth.

AI infrastructure is rising into that environment. The question is not whether AI can produce more output. It can. The question is whether more people can see themselves inside the output.

The fertility signal

Falling birth rates are often treated as a demographic statistic. They are also a confidence indicator. People have always made family decisions for private reasons:

career,

health,

relationships,

values,

money,

age,

housing,

fertility,

childcare, and

personal freedom.

What has changed is the level of perceived and real uncertainty.

The BBC’s recent reporting captured this well. One interviewee said the world does not feel like a nice place to bring children into. Another wanted more children but could not justify the cost. Those stories matter because birth rates do not fall through one cause. They fall when enough people, making private decisions, decide that the future feels too expensive, too uncertain, too pressured, or too difficult to commit to.

This is no longer a narrow UK story. The US is dealing with historically low fertility. France has crossed a symbolic line, with deaths exceeding births in 2025. Germany, Japan and South Korea are living with older population structures and weak replacement. China has also reversed direction sharply: after decades of population control, it is now trying to encourage births through childcare subsidies, expanded birth-related medical coverage and other family-support policies. The unresolved question is whether households will respond, given the cost of housing, education, childcare, job insecurity and the long shadow of earlier policy. India has a younger population, but faces a different question: can enough young people enter good work fast enough?

The common issue is confidence in the path ahead.

In some countries, that confidence shows up in births. In others, it shows up in youth employment. In others, it shows up in housing, delayed marriage, migration, political pressure or resistance to new infrastructure.

Demographics shape future demand, labour supply, pensions, tax bases, housing markets, healthcare systems, consumer behaviour and political mood. A low-birth society is also a high-dependency society. Fewer young people carry more of the future system: higher debt, older populations, housing shortages, infrastructure strain, climate costs, defence costs and the adjustment costs of AI.

Then they are told that technology will make the economy more productive. That may be true. It will matter more if young people can see a path into that productivity.

The missing route into work

The Milburn review is uncomfortable because it rejects the lazy explanation. The problem is not that young people do not want to work. The BBC reported that 84% of NEET young people surveyed want a job or training. That’s important. If people want work but cannot enter the system, the problem is more one of structure.

The entry point is failing. For instance, retail and hospitality vacancies have fallen sharply in the UK. Entry-level office work is being compressed by automation. Graduate recruitment is increasingly filtered through algorithms. Employers ask for experience before giving anyone the chance to get it. AI tools now do first drafts, summaries, analysis, customer support and coding scaffolds.

The old bargain was imperfect, but it was legible. You started with basic work. You learned by doing. You absorbed standards. You watched older people make decisions. You made small mistakes before you were trusted with larger ones. That bargain is weakening.

Companies can remove junior work before they have designed a replacement learning system. That may look efficient this quarter. It creates a capability problem later: no junior analysts, no trained managers, no institutional memory, no judgement pipeline, no leaders who learned the business from the bottom.

This is why the first rung matters. Starter jobs are not cheap labour. They are paid learning systems. When Boards remove them without replacing them, they reduce future competence.

The fix is not to preserve every old task. Some junior tasks should disappear. Nobody needs to spend three years formatting slides, reconciling low-value data, or copying numbers between systems because the organisation never fixed its processes. The fix is to replace accidental apprenticeship with deliberate apprenticeship.

That means real work, structured learning, AI fluency, human mentorship, rotations, accountability and progression. A company can automate weak work and still build strong people.

That is the real design challenge.

AI is changing the first job before the first job happens

The AI effect on young people starts before employment. It starts in applications. Young people now face automated screening, algorithmic ranking, AI-assisted tests, video interviews, psychometric filters and job descriptions that ask for skills the education system has not yet made standard.

At the same time, young people are told to use AI to compete. Write the CV with AI. Prepare for interviews with AI. Learn AI tools. Show AI literacy. Build a portfolio. Automate something. Create proof of work. That advice is not wrong. I have made parts of that argument myself.

But it creates a brutal asymmetry. The people with the least workplace context are being asked to master the tools reshaping work before they have learned how work operates. A graduate can use AI to write a better application. So can everyone else. Employers receive more polished applications, filter harder, and raise the bar again.

The system produces more output and less trust.

That is one reason the job search now feels so demoralising. The candidate sees effort. The employer sees volume. The algorithm sees patterns.

Nobody sees enough judgement.

This is where Boards should pay attention. Recruitment is not only a hiring function. It is the front door of the organisation. If the front door becomes opaque, expensive to navigate and biased toward those with coaching, connections and AI fluency, the organisation will start selecting for system-navigation rather than potential.

A better approach is available. Employers can publish clearer entry criteria, reduce performative application rounds, test real work rather than polish, and use AI to widen access rather than narrow it. They can create paid work trials, apprenticeships, internships and rotational programmes that develop judgement rather than merely screen for it. They can tell candidates what AI use is allowed and design entry routes for people without expensive coaching.

This is not charity. It is how companies rebuild access to talent.

Data centres and the local bargain

AI feels weightless in a browser. It is not weightless in the real world. It needs data centres, power, water, cooling, chips, transformers, grid capacity, land, planning consent and financing. It also needs local tolerance.

Data centres may fund useful technologies. They may support grid upgrades, new power projects, cooling innovation and industrial renewal. That upside matters. But communities do not experience “AI infrastructure” as a national productivity chart. They experience a planning application, a power connection, a water-use debate, construction traffic, heat, noise, limited local employment, and a site that may not feel connected to the lives around it.

Reuters has reported that data-centre opposition is spilling into French municipal politics, with candidates calling for moratoriums or greater transparency. That is a useful warning. The politics will not stay in France. It will spread because the build-out is physical.

People will ask a simple question:

What do we get?

That question will shape planning, regulation, permitting, grid access, taxation, community benefits, political campaigns and corporate reputation. The companies that answer badly will face delay. Delay has a cost.

The better answer is to make the bargain visible. Local power benefits. Local training. Local hiring. Local apprenticeships. Waste heat reuse where practical. Community funds. Transparent water use. Grid investment. Clear tax contribution. Open reporting. Community involvement before opposition hardens.

This is how infrastructure earns permission.

The participation problem

There is a common Boardroom instinct to place these issues in separate buckets:

• Birth rates go to government.

• NEETs go to education and welfare.

• AI goes to technology strategy.

• Data centres go to infrastructure and energy.

• Recruitment goes to HR.

• Planning objections go to public affairs.

That separation is tidy but it is no longer adequate.

The same participation problem runs through all of them: can people see a credible route into the future being described to them?

That question travels across borders. For an American household, it may be childcare, healthcare and housing costs. For a French household, it may be whether family formation still feels affordable and stable. For a German manufacturer, it may be whether the next cohort of skilled workers is large enough and trained enough. For Japan and South Korea, it may be whether ageing can be managed without asking too much of a shrinking younger generation. For India, it may be whether growth can turn education into good employment rather than frustration. For Europe more broadly, it may be whether youth unemployment and NEET rates remain manageable before they become political facts.

For a young person, participation means access to work and training. For a couple, it may mean a route into housing, childcare and family formation. For a community, it means seeing local benefit from local infrastructure. For employees, it means AI improves work rather than increasing pressure and reducing bargaining power. For Boards, it means productivity claims have to turn into measurable gains without weakening future talent, demand or trust.

This is the part that macro charts miss:

GDP can rise while confidence falls.

Productivity can improve while entry-level work disappears.

Infrastructure can expand while local permission weakens.

AI can be used everywhere while people trust the system around it less.

That is how a society ends up with growth and resentment at the same time.

The better path is to design participation into the growth model from the beginning.

How to square the circle

The gloom is not the point. The real point is that the route can be rebuilt:

AI can make companies more productive and help young people learn faster.

Data centres can support power systems and hard technologies.

Low birth rates can force societies to become more serious about housing, childcare, flexibility and career design.

Youth unemployment can be reduced when governments, employers and schools build practical routes into work.

The problem is not insoluble. It is badly organised.

There are seven practical moves that would change the direction.

1. Build paid entry routes into AI-era work

Every large employer using AI should have a serious answer to this question: how will young people learn here?

That answer cannot be a slogan about digital skills. It should include paid apprenticeships, graduate rotations, entry-level roles redesigned around AI, and managers rewarded for developing junior staff. A junior employee should learn how to use AI, check AI, challenge AI and do the underlying work well enough to know when AI is wrong.

That is the new apprenticeship.

2. Replace weak tasks with stronger learning

Some old entry-level tasks are disappearing. Good. But the learning they carried cannot disappear with them.

So, if AI:

• now drafts the first version of a report, the junior should learn how to test assumptions, trace sources, check logic, rewrite for a decision-maker, and understand the commercial consequence;

• classifies invoices, the junior should learn controls, exceptions, fraud patterns, data quality and process design;

• writes code, the junior should learn architecture, testing, security, maintainability and user need.

The task changes and training must become more intentional.

3. Make AI literacy universal and practical

AI literacy should be treated like spreadsheet literacy became in previous decades. But it must go beyond prompts.

People need to understand what AI is good at, where it fails, how to verify output, how to protect confidential data, when to use a small model, when to use a frontier model, when a human must decide, and how to keep a record.

A workforce that uses AI without understanding verification will create hidden risk. A workforce that understands AI can improve productivity and judgement at the same time.

4. Link infrastructure to local benefit

Data centres and AI infrastructure need local bargains. Companies and governments should publish what a project gives locally, not just nationally:

How many jobs?

What training?

What grid investment?

What community benefit?

What water use?

What heat reuse?

What power source?

What tax contribution?

What protections?

The answer should be specific enough for local people to judge. If the answer is weak, the project should be improved.

5. Measure verified productivity, not activity

Boards should stop counting AI activity as AI value. Prompts sent are not value. Documents generated are not value. Tools deployed are not value. Agents launched are not value.

The better measures are harder and more useful: output per hour after review, unit cost per verified outcome, error rates, rework, cycle time after human approval, revenue per employee, customer satisfaction after automation, cost of control, cost of failure, and vendor dependency.

The company has proved AI productivity only when old work disappears, the new cost base is counted, quality holds, risk stays within appetite, and the gain reaches the P&L.

6. Give people a stake in the productivity gain

AI will face less resistance if people see some upside. That can happen through wages, training, shorter repetitive work, better career progression, equity participation, profit-sharing, lower prices, improved services, or better public outcomes.

Not every gain has to go to employees. Shareholders took risk and capital matters. But if all gains go to capital and all adjustment goes to workers, the bargain will weaken.

People support systems they can join.

7. Treat family formation as economic infrastructure

Many people will not have children. That is a valid choice. But people who want children should not be blocked by housing, childcare, unstable work and hostile career structures.

Companies cannot fix national fertility alone. They can still help through flexible work that does not punish parents, predictable scheduling, childcare support where feasible, career paths that survive parental leave, men expected to parent, returnships, manager training, and promotion systems that do not treat family life as lack of ambition.

Governments have larger levers: housing, childcare, tax, education, healthcare, transport, planning and migration. But employers shape the daily reality. A society that wants future workers, consumers and taxpayers should make it easier for people who want families to have them.

The Board questions that matter now

Boards do not need to solve fertility policy or youth unemployment alone. They do need to stop pretending these are externalities with no relevance to strategy.

A useful Board discussion should cover ten points:

1. Which junior roles are being removed, compressed or redesigned by AI?

2. What replaces the learning those roles used to provide?

3. Does the hiring process identify potential, or does it reward candidates who know how to pass automated filters?

4. Which AI gains have reached the P&L after compute, software, integration, review, security and rework?

5. Are employees using AI because it improves work, or because they fear being left behind?

6. Where does the company create local cost or remote benefit, especially around infrastructure, offices, energy, data centres, logistics and automation?

7. What evidence would make the Board reduce its AI productivity assumptions?

8. How does the company share productivity gains with workers, customers or communities?

9. Which skills must be protected because the company will need them in 5 or 10 years?

10. Does the company assume future consumers, taxpayers, workers and families will exist in sufficient numbers and with sufficient confidence to support its plans?

The last point sounds rather broad: it is not. Companies depend on social reproduction in the literal and economic sense. They need new workers, new customers, new managers, new households and new taxpayers. A society that cannot give young people a route into work and family formation is eating future demand.

The hard truth for AI strategy

AI strategy cannot stay inside the productivity box. That’s because the productivity box is too small.

AI impacts the first job. It interacts with training and who captures gains. It affects where infrastructure gets built and energy demand. It affects trust in work and whether young people think the future is something they can enter or something that happens to them.

A good AI strategy must answer more than how to automate. It must answer who:

• learns,

• gains,

• verifies,

• is displaced,

• is hired,

• is trained,

• carries the cost,

• sees the benefit, and

• gives permission.

If those questions sound political, that is because they are. But they are also commercial. Legitimacy is now part of execution. Recruitment quality is part of AI strategy. Training is part of resilience. Local acceptance is part of infrastructure finance. Demographics are part of long-term demand.

The companies that understand this will still use AI aggressively. They will automate. They will reduce waste. They will build or buy compute. They will use agents. They will redesign workflows.

They will also preserve the routes that make growth believable.

The weekend takeaway

The future is being built, data centres are rising and AI systems are spreading. Compute is being priced, traded and allocated. Boards are being told that productivity will improve, costs will fall and growth will follow.

Some of that will be true.

When young people cannot get work, the first rung disappears. When families cannot afford children, demographic confidence falls. When communities host infrastructure without visible benefit, permission weakens. When employees use AI while trusting it less, adoption numbers mislead management. When Boards count productivity but ignore participation, they mistake output for resilience.

But this can be fixed:

Build paid entry routes.

Redesign junior work.

Teach AI literacy properly.

Connect infrastructure to local benefit.

Measure verified productivity.

Share some of the gain.

Make work and family life more compatible.

The useful question is simple:

If they can, AI becomes part of renewal.

But, if they cannot, the growth story weakens and that not because the technology failed, it is because the bargain fails.

Leave a comment

🧠 Techie Corner for the Non-Techies

1. NEET

NEET means not in education, employment or training. It is one of the clearest measures of whether young people are connected to work or learning.

A high NEET rate matters because it can create long-term effects: lower earnings, weaker confidence, poorer health, higher welfare costs and reduced future tax receipts. For companies, the NEET issue matters because it reduces the future talent pool.

2. Fertility rate

The total fertility rate estimates the average number of children a woman would have over her lifetime if current birth rates continued.

A fertility rate below 2.1 is usually below replacement level in advanced economies, although migration, mortality and age structure also matter. This matters for Boards because demographics affect labour supply, consumer demand, pensions, healthcare costs, tax capacity and political priorities.

3. Entry-level compression

Entry-level compression happens when technology removes or reduces the basic tasks that used to train junior workers. Examples include drafting first versions, summarising documents, preparing reports, running simple analysis, answering routine customer queries and checking standard forms.

These tasks may be boring but they also teach context, judgement and standards. When AI removes them, companies need new training routes.

4. Verification debt

Verification debt is the checking burden created by machine-generated output. AI can produce more drafts, summaries, code, analysis and recommendations than an organisation can easily review.

The output may look polished and that makes mistakes harder to spot. A company has gained productivity only when useful output rises after review, correction, risk control and accountability are included.

5. Data-centre externalities

An externality is a cost or benefit that falls on someone outside the main transaction. Data centres can create local externalities: power demand, grid pressure, water use, construction disruption, land use, noise, heat and limited local employment.

The AI buyer may see compute capacity, yet, the community sees infrastructure pressure. That gap drives political risk.

6. Licence to operate

Licence to operate means the practical permission a company needs from regulators, communities, customers, employees and politicians. It is wider than legal permission.

A company can have formal approval and still face resistance that slows projects, raises costs or damages trust. AI infrastructure will increasingly depend on licence to operate because the costs are physical and local.

7. Participation

Participation means people can enter and benefit from the system. For young people, participation means access to work and training. For families, it means confidence to form households and raise children. For communities, it means seeing local benefit from local infrastructure. For employees, it means AI improves work rather than simply increasing pressure.

Productivity without participation becomes politically fragile.

8. Deliberate apprenticeship

Deliberate apprenticeship means replacing accidental learning with designed learning. The old model relied on junior people learning by doing basic tasks. AI will remove many of those tasks.

The new model must teach people through structured work, AI-assisted practice, human review, rotations, feedback and accountability. That is how companies keep developing judgement while still using AI.

📚 Further reading

• BBC News: “It’s not a nice world to bring children into”: Births fall to the lowest level in 50 years (2026)
  Useful for the UK fertility and confidence hook. Reports 585,000 live births in England and Wales in 2025 and a fertility rate just under 1.4.

• Office for National Statistics: Births in England and Wales: 2025 (2026)
  Primary UK source for births, fertility and demographic data.

• BBC News: Opportunities shrinking for too many young people, says major report on “lost generation” (2026)
  Useful for the UK youth-employment hook. Reports more than 1 million young people not in education, employment or training and Alan Milburn’s warning about a lost generation.

• Alan Milburn / UK government: Young People and Work review materials (2025-2026)
  Policy context for the government-commissioned review into rising numbers of young people who are NEET.

• CDC / National Center for Health Statistics: Births: Provisional Data for 2025 (United States, 2026)
  Primary source for US provisional birth data. Reports about 3.6 million births in 2025 and a general fertility rate of 53.1 births per 1,000 women aged 15-44.

• USAFacts: How have US fertility and birth rates changed over time? (2026)
  Useful US public-data source. Gives a clear non-partisan presentation of the long-term decline in US fertility and birth rates.

• INSEE: Demographic report 2025 (France, 2026)
  Primary French source. Reports 645,000 births in 2025, fertility at 1.56 children per woman, and deaths exceeding births.

• le Monde: France records more deaths than births in 2025, a first since World War II (2026)
  Le Monde reports 651,000 deaths and 645,000 births in 2025, with deaths exceeding births for the first time since World War II.

• Destatis: Decline in fertility rate slowed significantly in 2024 (Germany, 2025)
  Primary German source. Reports Germany’s fertility rate at 1.35 children per woman in 2024.

• Statistics Bureau of Japan: News Bulletin on the number of children (2025)
  Primary Japanese source. Reports Japan’s child population at record lows and continuing long-term decline.

• OECD: Economic Surveys Korea 2024, responding to population decline (2024)
  Institutional source on South Korea’s demographic position. Describes Korea’s fertility rate as the world’s lowest, around 0.7 in 2023.

• ILO and Institute for Human Development: India Employment Report 2024 (2024)
  Useful source on India’s youth employment challenge. Highlights educated youth unemployment and the large share of young people, particularly young women, outside education, employment or training.

• Eurostat: Statistics on young people neither in employment nor in education or training (2026)
  Primary EU source for NEET data.

• Eurostat: Unemployment statistics (March 2026)
  Primary EU source for youth unemployment.

• Reuters: Backlash against data centres is spilling into French municipal election races (2026)
  Useful for the infrastructure-legitimacy angle. Shows data centres moving from technology infrastructure into local politics.

• Not Boring by Packy McCormick: Thank God For Data Centers (2026)
  Useful counterpoint. Argues that AI data centres may act as buyers of capability, funding technologies such as nuclear, geothermal, grids, batteries and other hard technologies before traditional markets would.

• Dr Henri Winand: AI & the Missing First Rung of the Career Ladder (2025)
  Earlier ‘Something for the Weekend’ article arguing that starter jobs are investments in resilience, not cheap labour.

• Dr Henri Winand: The Legitimacy Risk No One Is Pricing (2026)
  Earlier ‘Something for the Weekend’ article on the risk that AI capability scales faster than public acceptance.

• Dr Henri Winand: The Productivity Miracle Someone Else Pays For (2026)
  Earlier ‘Something for the Weekend’ article on AI productivity, asset prices, debt relief and distribution.

• Dr Henri Winand: Rent the intelligence. Own the proof. (2026)
  Earlier ‘Something for the Weekend’ on compute, local AI, smaller models and audit trails.

• China State Council / Xinhua: China to offer nationwide childcare subsidies (2025)
  Primary Chinese government-linked source. Reports China’s nationwide childcare subsidy programme from 2025, offering 3,600 yuan per year for each child under three.

• Reuters: No “out-of-pocket expenses” for childbirth in China from 2026 (2025)
  Reports China’s plan to expand birth-related medical coverage, including prenatal checks, as part of wider efforts to encourage births.

• National Bureau of Statistics of China / 2025 population data: Statistical Communiqué of the People’s Republic of China on the 2025 National Economic and Social Development (2026)
  Primary source to verify China’s population and birth trend. Use with care, as English-language summaries and third-party reporting may differ in timing and presentation.

Last Week on Something for the Weekend

Last week’s piece looked at the operating bargain under enterprise AI. The argument was that compute may become priced, traded and allocated, but serious AI decisions still need control and proof.

Cloud AI gives access to powerful models, but it is metered. Local AI gives more control, but it creates capital cost and utilisation risk. Small language models may suit narrow, private workflows better than large models. Audit trails matter because serious decisions need a record.

The useful question was:

Where did the answer run, what did it cost, and can we prove why it was given?

That piece looked at the machine-work layer. This week looks at the human layer underneath it: whether young people, families, workers and communities still see a route into the future that AI is helping to build.

A trader wants to hedge the price of GPU capacity, and a lawyer wants to run privileged client documents through an AI system without sending them to someone else’s cloud.

But the CFO wants to know whether the new AI agent is a subscription, a cloud meter, or a capital asset: “am I dealing with a fixed, variable or semi-variable cost model, how do I forecast costs versus revenues and productivity?”.

Of course, the software team wants access to the best frontier model for a hard coding problem, yet, the compliance officer wants something narrower: an answer linked to the source, the policy, the model version, the prompt, and the human approval.

Some are buying rented intelligence. Some are buying owned machine capacity. Some are buying something more demanding: a decision that carries proof.

The AI market is starting to split along those lines.

One part of the market wants compute to become tradable, benchmarked, and hedgeable. Larry Fink has been talking about futures markets for computing power. CME and Silicon Data have announced plans for compute futures linked to GPU rental prices, pending regulatory review.

Another part of the market is moving in the opposite direction. People are buying local machines because they want AI that runs close to their data, their workflows, and their clients. 9to5Mac recently argued that demand for local AI could create a new business model for Apple, driven partly by buyers using Mac minis and Mac Studios as local or headless AI machines.

Both directions can be true.

Commodity compute will matter because AI needs scale. Local compute will matter because some data should stay controlled. Smaller models will matter because some tasks need fidelity more than breadth. Audit trails will matter because serious decisions need a record.

These are the next AI questions:

• Where did the answer run?

• What did it cost?

• Can you prove why and how this answer was given?

The Boardroom point

The next AI decision is about where the intelligence runs, how it is paid for, and whether the answer can be reconstructed later:

Cloud AI is metered operating expenditure.

Local AI is capital expenditure with utilisation risk.

Large language models buy breadth.

Small language models can buy better control.

Audit trails buy accountability.

This is a direct extension of a point I made recently in an earlier Something for the Weekend article:

The digital employee was never $200 a month. The subscription was the entry price and the full cost included tokens, inference, context, API calls, retries, cloud infrastructure, integration, verification, and risk. The human employee appeared expensive. The digital employee was metered.

That was the cost-accounting problem.

The next problem is somewhat more strategic:

• Once AI work is metered, companies need to decide where that work should run. Some machine work belongs in the cloud because it needs frontier models, large context windows, burst capacity, and constant upgrades.

• Some machine work belongs locally because it is confidential, narrow, repetitive, or sensitive enough to require direct control.

• Some machine work should only happen if the answer comes with a record of the data, model, source material, rules, and approvals behind it.

That is where AI moves from productivity software into governance.

Boards do not need to approve every AI tool but they must understand (i) the cost structure and (ii) control structure behind machine work.

Why this is happening now

The AI build-out has become too large to hide inside software budgets.

BloombergNEF reported in March that capital expenditure by the 14 largest publicly owned data-centre operators globally was expected to approach $750 billion in 2026, compared with a little less than $450 billion in 2025. It also estimated that 23.1GW of data-centre IT capacity was under construction globally at the end of September 2025, including 15.9GW in the US alone.

That is physical infrastructure on a grand scale which needs land, power, chips, memory, cooling, financing, contracts, and customers. It also needs prices.

That is why Larry Fink’s comment matters. Bloomberg reported this month that the BlackRock CEO expects demand for computing power to become large enough for traders eventually to bet on futures for compute. His argument sits inside a broader scarcity problem around compute, chips, and memory.

The market is already trying to build the instruments. CME Group and Silicon Data announced on 12 May 2026 that they plan to launch compute futures later this year, pending regulatory review. The contracts are expected to use Silicon Data’s daily GPU benchmarks for on-demand rental rates and are aimed at AI builders, cloud providers, financial institutions, and traders managing compute price volatility.

The logic is clear enough:

When an input becomes large, scarce, and volatile, finance tries to price it, trade it, and hedge it.

Compute now fits that description.

The harder part is that compute is not a clean commodity. A GPU-hour depends on chip type, memory, location, latency, power reliability, cooling, software stack, workload, data rules, and service levels. A unit of compute in Virginia may have a different value from a unit of compute in Singapore. A training cluster has a different use from a local inference machine. A frontier model call has a different cost profile from a small model classifying invoices. And the local cost of electricity must be factored in.

A futures market can price a part of this but it cannot make every kind of compute interchangeable.

From metered labour to metered intelligence

The first budgeting mistake was confusing access with work:

A subscription gives access to a system. It does not guarantee unlimited economic output.

The real cost depends on what the AI does. A short answer may cost very little. A long document review costs more. An agent that reads files, calls tools, writes code, tests it, fails, retries, and produces a long report costs more again.

The visible price is the subscription and the operational price is the meter.

That meter can sit in several places. It can sit with the AI vendor. It can sit with the cloud provider. It can sit inside the company’s own hardware. It can also sit across a hybrid system where the company owns some local capacity and rents extra capacity when it needs more.

Each version changes the economics.

Cloud AI converts machine work into variable cost. Local AI converts machine work into capital cost. A cloud agent can become expensive through use. A local machine can become expensive through underuse.

The CFO has to understand both.

A company can move work from people to machines and still fail to create economic value if the machine work creates uncontrolled usage, duplicate workflows, verification debt, or new legal exposure.

That is why AI cost accounting needs to move beyond licences and seats.

What did the system do? Was it correct? Was it used? Did it remove old work, or did it add another layer? Did it reduce cost, reduce risk, increase revenue, or move the bottleneck somewhere else?

Those questions become more important as compute becomes more expensive, more visible, and more tradable.

Share Something For The Weekend

The rented brain

Cloud AI is the ‘rented brain’:

It has obvious advantages. It gives users access to frontier models without buying the infrastructure. It scales when demand spikes. It lets smaller companies use tools they could not build themselves. It gives teams access to models that improve quickly.

For many tasks, cloud AI is the right answer.

Coding assistants, research tools, drafting tools, general productivity systems, customer-support workflows, and exploratory analysis often benefit from the best available model and flexible usage.

The weakness sits in the cost and control model.

Cloud AI depends on the provider’s infrastructure economics. Heavy usage consumes real compute. Long context windows consume more tokens. Agentic workflows consume more steps. Tool calls, retries, and model upgrades all matter.

Rate limits are the visible edge of that cost structure.

Reuters reported that SpaceX agreed to lease its Colossus 1 AI data centre in Memphis to Anthropic, giving Anthropic access to more than 220,000 Nvidia processors and 300MW of new capacity within a month. Reuters also reported that the deal would help Anthropic raise limits for Claude Code and API usage.

That is the economic signal:

Usage limits rise when capacity rises. They tighten when capacity is scarce.

The consumer sees a chatbot. The enterprise sees a tool. The supplier sees inference cost.

The CFO should see a meter.

The owned machine

Local AI reverses the economics:

The cost becomes visible upfront. The company buys the hardware, provisions the memory, configures the stack, manages security, accepts constraints, and carries the risk that the hardware becomes underused or obsolete.

That can look unattractive compared with a cloud subscription.

For some work, visible cost is useful.

A local system can keep sensitive data inside the organisation. It can run against a defined corpus. It can be versioned. It can be frozen. It can be tested. It can produce a more repeatable process. It can also reduce dependence on a provider’s pricing, access rules, rate limits, and policy changes.

This is pretty important for law firms, family offices, insurers, banks, healthcare providers, boards, research teams, and anyone working with confidential or regulated material.

9to5Mac’s local-AI piece is useful because it captures a real behavioural signal. Some buyers are not simply chasing faster hardware. They want local capacity because they do not want every sensitive task to become a cloud call.

Of course, local AI does not remove cost, but it certainly changes the type of cost.

Cloud AI creates usage risk: cloud AI can run away through consumption.

Local AI creates capital risk: local AI can disappoint through poor utilisation.

The right answer depends on the workload. So, a company running occasional, broad, high-complexity work may prefer cloud models. A company running frequent, narrow, confidential inference may prefer owned capacity. A company doing both needs routing rules.

Which work goes where becomes the practical issue to wrestle with.

SLMs and LLMs

The local versus cloud debate becomes more useful once model size is added.

Large language models (LLMs) are generalists. They are useful where breadth, language fluency, flexible reasoning, coding, drafting, summarising, and multi-domain capability matter.

That breadth is valuable but it is also more expensive.

Small language models (SLMs) have a different use case. They know less. They can also be easier to constrain, cheaper to run, easier to deploy locally, and better suited to narrow workflows.

For some tasks, knowing less is acceptable because the system should only operate inside a defined boundary.

A small model may be suitable for document classification, invoice matching, policy checks, structured extraction, internal knowledge search, contract comparison, claims routing, meeting summaries, and first-pass reviews against a defined rule set.

The useful test is simple:

• Does the task need broad reasoning across the open world, or does it need controlled reasoning across a defined corpus?

If the task needs breadth, use a larger model. If the task needs repeatability, privacy, and source discipline, a smaller model may be enough.

SLMs are not a downgrade when the task is narrow: they are a governance choice.

That does not make them safe by default. A badly designed small-model system can still produce wrong answers. It can still miss context. It can still propagate errors. It can still create false confidence.

The case for SLMs is about fit, not some sort of ‘magical attributes’.

A smaller model doing a narrow job well may be better than a larger model doing the same job expensively, inconsistently, or outside the organisation’s control.

Fidelity is not fluency

AI has made fluent output cheap.

That creates a management problem because a fluent answer can still be wrong.

A polished answer can cite the wrong source. A confident answer can miss the exception. A persuasive answer can rely on the wrong version of a document.

For serious work, fidelity matters more than style.

Fidelity means the answer is faithful to the right material: the right policy, the right contract, the right data room, the right jurisdiction, the right model version, the right approval path, and the right constraint.

That is why model choice cannot be separated from workflow design.

A frontier LLM may give the most fluent answer. A smaller local model may give a more controlled answer if the workflow restricts the corpus, retrieves the relevant passages, applies defined rules, and keeps a record.

The difference is important for legal work, audit, insurance, finance, healthcare, public services, and board processes:

A board paper that reads well but uses the wrong assumption is dangerous.

A contract summary that sounds authoritative but misses the operative clause is dangerous.

A claims recommendation that cannot show which policy wording it relied on is dangerous.

The answer needs a source record.

The decision receipt

A serious AI decision should carry its own ‘receipt’.

That receipt should show what data was used, which corpus was searched, which embedding index was queried, which passages were retrieved, which model and version ran, which prompt template applied, which rule or policy constraint was used, what the model output was, who reviewed it, who changed it, and who approved it.

This is how AI moves into serious work: the output alone is too thin.

A company needs to know how the answer was produced. It needs to know which facts were used, which facts were ignored, which rules were applied, and which person accepted responsibility.

Recent research on LLM audit trails makes this point directly. The authors describe audit trails as chronological, tamper-evident, context-rich records that link technical provenance, such as models, data, training, evaluation, deployment, and monitoring, with governance records, such as approvals, waivers, and attestations.

That is the right direction for enterprise AI.

An AI answer without a record is hard to challenge, reproduce, or defend.

For low-risk work, that may be acceptable, but for really consequential work, it will not, and should not, be acceptable.

If an AI system rejects a customer request, the company should know why. If an AI system flags a transaction, the company should know why. If an AI system summarises a contract, the company should know which clauses it relied on. If an AI system drafts a board paper, the company should know which assumptions it used.

“The model said so” is not an audit trail.

The compute portfolio

The future enterprise AI stack will be a portfolio of compute choices where some work will use:

• commodity cloud compute because it needs scale;

• frontier cloud models because it needs breadth;

• local compute because the data is sensitive;

• smaller models because the task is narrow;

• audit trails because the decision must survive scrutiny.

The company that manages this well will route work by (i) cost, (ii) risk, (iii) fidelity, and (iv) proof:

A broad research question may go to a frontier model.

A confidential contract comparison may stay local.

A repetitive compliance check may use an SLM.

A high-volume workflow may justify owned capacity.

A board-level decision may require an audit trail and human approval.

Even it that sounds less exciting than buying an AI tool, it is intrinsically more useful.

The winners will not be the companies with the most AI experiments. They will be the companies that know which work to run where, what it costs, and how to prove the result.

The CFO version

For the CFO, this is a cost-structure problem:

• Cloud AI is variable cost: a cloud agent can become expensive through usage.

• Local AI is capital cost: a local machine can become expensive through underuse.

• Hybrid AI is both: a frontier model can be expensive because it is powerful.

• A smaller model can be cheap and still inadequate if it is used for the wrong task.

The cheapest output is not always the cheapest work.

A cheap answer that creates verification debt may be expensive. A costly answer that prevents legal, regulatory, operational, or reputational error may be good value.

The CFO should ask for the fully loaded cost of verified machine work.

That includes model access, tokens, inference, context, cloud, hardware, depreciation, energy, integration, security, monitoring, human review, errors, vendor risk, downtime, and governance.

The old software budget was mostly about licences, implementation, seats, and renewals.

The AI budget is about machine work, and that needs cost accounting.

The CIO version

For the CIO, this is an architecture problem since the company needs routing rules:

• Which workloads require frontier models?

• Which workloads require local control?

• Which workloads can use smaller models?

• Which workflows need a retained evidence record?

• Which outputs need human review?

• Which systems create vendor concentration risk?

• Which systems would fail if a provider changed pricing, model access, rate limits, or usage policies?

• Which internal data can be sent to external systems, and which data should stay inside the organisation?

The CIO needs to connect model choice, data governance, security, cost, workflow design, and audit. This is not a procurement optimisation exercise, it is more akin to design of core business operations.

The Board version

For the board, this is an accountability problem. Boards do not need a technical debate about every model, vendor, or parameter count. They need clear answers about exposure.

Five questions matter:

1. Where is AI already being used in production?

2. Which workloads are cloud-based, local, hybrid, or embedded inside vendor software?

3. Which costs are fixed, variable, capitalised, or hidden inside other budgets?

4. Which AI outputs require human approval?

5. Which AI decisions can be reconstructed later?

That last question will matter more over time. If the company cannot show how a decision was made, it may struggle to defend the decision. And that applies to customers, employees, regulators, auditors, boards, insurers, and courts.

The issue is accountability.

Leave a comment

The weekend takeaway

The metered digital employee was the first warning, and the next warning is larger:

Once compute becomes priced, traded, rationed, and allocated, every organisation needs a compute strategy.

Some companies will rent intelligence from the cloud and some will own local capacity. Some will use smaller models for narrow workflows, yet others will need audit trails for decisions that can be challenged later.

Most will need a mix: the most expensive mistake will be treating all AI work the same since some:

• tasks need scale;

• need privacy;

• need fidelity, and

• some need proof.

The valuable AI system may not be the one with the largest model. It may be the one that gives the right answer, from the right evidence, inside the right workflow, with the record attached.

So, compute may become tradable, but trust has to be earned one decision at a time.

The useful boardroom question is direct:

🧠 Techie Corner for the Non-Techies

1. Compute

Compute is the processing power used to run digital work.

In AI, compute usually means the hardware and infrastructure needed to train or run models. That includes chips, memory, servers, power, cooling, networks, and software.

Training is the expensive process of building or improving a model.

Inference is the repeated process of using the model to answer a prompt or complete a task.

Most companies will not train frontier models but many will pay for inference.

That is where the meter appears.

2. GPU

A GPU is a graphics processing unit. It was originally designed for graphics, but it is also good at the type of calculations used in AI. Modern AI infrastructure depends heavily on GPUs and other accelerators.

Not all GPU clusters are equal. Chip type, memory, interconnect, power supply, cooling, and location all affect the value of the compute.

That is why one GPU-hour is not automatically the same as another.

3. Token

A token is a unit of text or data processed by an AI model.

Short prompts use fewer tokens. Long documents, long conversations, codebases, and agentic workflows use many more.

Tokens matter because many AI systems are priced, limited, or managed according to how much input and output the model processes. A model that reads more context and produces mode completed answers usually consumes more tokens which can increase cost.

4. Context window

The context window is the amount of information an AI model can consider at one time. A larger context window lets the model read more documents, code, or conversation history. It can also cost more because the model must process more material. That is why giving the AI every available document is often inefficient.

Better systems retrieve the relevant material first, then ask the model to work from that smaller set.

5. LLM

An LLM is a large language model. It is designed to handle broad language tasks: drafting, summarising, coding, reasoning, translation, analysis, and conversation.

LLMs are useful because they are flexible, but they are also more expensive to run compared with smaller, narrower models.

For many tasks, a frontier LLM is the right tool but for some, it is simply an overkill.

6. SLM

An SLM is a small language model. It has fewer parameters and usually less general capability than a large model. That does not make it useless.

An SLM can be useful where the task is narrow, the data are private, the workflow is repeatable, and the organisation wants lower cost, local deployment, or tighter control.

The key is task fit.

A small model doing a narrow job well may be better than a large model doing a narrow job expensively.

7. Local AI

Local AI means running the model on hardware controlled by the user or organisation, rather than sending the task to an external cloud service. That hardware could be a laptop, workstation, local server, private cluster, or on-premise data centre.

Local AI can help with privacy, latency, repeatability, and cost control. It also creates responsibilities: hardware cost, security, maintenance, model management, electricity, cooling, and upgrades.

8. Cloud AI

Cloud AI means using AI models and infrastructure provided over the internet by a vendor. It is attractive because it gives fast access to powerful models without buying the infrastructure. It is also metered.

Heavy use can increase cost, and access may depend on the vendor’s rate limits, pricing, policies, and capacity. Cloud AI is flexible but, it certainly is not costless.

9. Embedding

An embedding is a numerical representation of text, images, or other data. Embeddings help AI systems find related material.

For example, a company might turn its internal documents into embeddings so an AI system can search for relevant passages before answering a question.

Embeddings are useful because they help connect a question to the right evidence. They also create governance questions. The organisation needs to know which documents were embedded, when, from which source, and under which access rules.

10. Audit trail

An audit trail is the record of what happened.

In AI, a useful audit trail may include the model used, model version, prompt, retrieved documents, source passages, rules applied, human approvals, overrides, timestamps, and final output. This matters because an AI answer without a record is hard to challenge, reproduce, or defend.

‘A serious decision should carry its own receipt’.

Subscribe now

📚 Further reading

• BloombergNEF: AI Data Center Build Advances at Full Speed: Five Things to Know (2026)
  BNEF reports that capex by the 14 largest publicly owned data-centre operators globally is expected to approach $750 billion in 2026, and that 23.1GW of data-centre IT capacity was under construction globally at the end of September 2025.

• Bloomberg News: Larry Fink Predicts Birth of Futures Market for Computing Power (2026)
  Reports Fink’s view that compute demand is large enough to support futures markets, with shortages in compute, chips, and memory as part of the argument.

• CME Group and Silicon Data: CME Group and Silicon Data Partner to Launch First Compute Futures (2026)
  Primary announcement of planned compute futures based on daily GPU rental-rate benchmarks, pending regulatory review.

• 9to5Mac: The demand for local AI could shape a new business model for Apple (2026)
  A useful signal on local AI demand, Apple Silicon, high-memory Macs, and the tension between local control and cloud-hosted compute.

• Reuters: Anthropic strikes SpaceX data center deal as it plows ahead on AI coding (2026)
  Reports the SpaceX and Anthropic data-centre arrangement, including Colossus 1, more than 220,000 Nvidia processors, and 300MW of new capacity.

• Victor Ojewale, Harini Suresh and Suresh Venkatasubramanian: Audit Trails for Accountability in Large Language Models (2026)
  Academic treatment of LLM audit trails as context-rich, tamper-evident records linking model, data, deployment, monitoring, and governance events.

• Dr Henri Winand: The Digital Employee Was Never $200 a Month (2026)
  Last month’s SftW on why AI labour should be treated as metered machine work, rather than fixed-price employment. The current piece builds from that cost-accounting point into compute strategy and accountable decisioning.

Last Week on Something for the Weekend

Last week’s piece looked at the political and economic temptation sitting behind the AI productivity story.

The argument was that AI may raise productivity, lower unit labour cost pressure and give central banks more room to cut rates, much as Alan Greenspan believed personal computing had done in the late 1990s. The risk is that markets, governments and companies start spending the productivity gain before it has been measured.

That matters because AI productivity can be real and still unevenly distributed. Asset owners, large technology platforms and highly skilled AI users may capture the upside first, while salary earners, younger workers, local communities and non-asset owners carry more of the adjustment.

The core test was simple:

• Did output rise?

• Did the full cost fall?

• Did the gain spread beyond the owners of the machine?

That piece looked at the macro bargain around AI productivity.

This week looks at the operating bargain underneath it: where the AI runs, who controls the meter, and whether the decision can be proved after the event.

A new Fed chair walks into office with the most convenient economic story available:

AI may let America cut rates without paying the inflation bill.

That is the line that caught my eye in Morning Brew’s note on Kevin Warsh replacing Jerome Powell.

Powell leaves after an extraordinary 8-year period: early Trump-era rate rises, the Covid rescue, massive bond purchases, emergency support, a rapid labour-market recovery, then the worst inflation surge in decades and the painful correction that followed.

Warsh inherits the other side of that story.

He has said he favours lower rates, but the Fed is divided, inflation remains above target, and the committee is unlikely to move simply because the new chair wants it to.

The interesting part was not the personnel change.

It was the historical comparison.

Morning Brew pointed back to Alan Greenspan. In the 1990s, Greenspan resisted calls to raise rates because he believed the personal-computing boom had raised productivity. The economy could run faster without creating the same inflation pressure.

He was broadly right on that narrow call.

Warsh appears to see AI in similar terms.

That is what made me stop.

AI may well raise productivity. I have been writing about that from several angles: the digital employee, the superworker, the shrinking need for some corporate roles, the cost of machine work, and the legitimacy risk around who benefits and who pays.

There is a difference between productivity arriving and productivity being spent in advance.

That is this week’s question.

Who gets the AI productivity gain before markets, policymakers and companies start using it?

The Boardroom point

AI productivity is being priced before it has been proven:

Markets are pricing it.

Vendors are pricing it.

Governments are relying on it.

Boards are building cost plans around it.

Central bankers are starting to discuss it as a macroeconomic variable.

Some of this is rational. AI already changes real work. It writes, searches, codes, summarises, translates, drafts, checks and reasons across documents. Used well, it removes hours from repetitive cognitive work.

Some of it is premature:

A productivity boom is not the same as a usage boom.

A productivity boom is not the same as a capex boom.

A productivity boom is not the same as a valuation boom.

A company can deploy AI everywhere and still fail to reduce its full cost base. A worker can produce more and capture less. A government can get higher nominal growth while households experience weaker real purchasing power.

Boards should separate four things that are often blended together:

1. Productivity.

2. Asset prices.

3. Debt relief.

4. Distribution.

AI can improve the first while intensifying the other 3.

The useful Board question is direct:

Why this is happening now

The Warsh-Greenspan comparison is not some sort of random observation.

Greenspan’s argument had an economic mechanism behind it. Higher labour productivity can lower unit labour costs. If wages rise 4% and productivity rises 3%, the labour cost embedded in each unit of output rises by roughly 1%.

That gives the economy more room.

Greenspan made that argument in 1999. He said stronger labour productivity had lowered unit labour cost growth and reduced inflation pressure. He linked that change to the spread of information technology across business processes.

The data later supported the broad claim. The late 1990s did see a real productivity acceleration. Business investment in computing, software and networks changed how firms operated. Inflation stayed more contained than many models expected.

That historical memory now sits inside the AI debate.

Warsh inherits an economy with a much worse fiscal backdrop. Public debt is higher. Interest costs matter more. Political pressure on the Fed is more open. AI investment is already enormous. Equity markets have capitalised a large part of the promised gain.

The New York Times piece on Warsh makes the comparison explicit: Greenspan resisted calls to raise rates because he believed the US was experiencing a productivity boom from personal computing; Warsh has said AI may create a similar boom today.

That is the official version of the seduction.

There is a harder version.

When debt is high, productivity optimism becomes useful. It allows policymakers to believe that growth can carry what fiscal policy cannot fix. It allows markets to believe that valuations are justified. It allows companies to believe that cost savings will arrive. It allows voters to be told that the pain is transitional.

That makes the provision of evidence matter more.

The Greenspan lesson

Greenspan’s late-1990s call had two parts:

The first part was right: technology raised productivity and reduced inflation pressure for a period.

The second part was more fragile: easy financial conditions and technology optimism helped produce excess. Productivity gains were real. The capitalisation of those gains went too far.

This distinction is important.

Technology can raise the economy’s speed limit. It can also raise expectations faster than output. Asset prices can move before wages. Capex can move before cash flow. Political pressure can move before the data.

That is the Fed’s problem.

Central banks have to decide in real time, yet productivity data arrive late:

Revisions matter.

Business surveys overstate and understate.

Hype contaminates judgement.

Markets try to front-run the conclusion.

Chicago Fed President Austan Goolsbee has already warned about this. Rising productivity can restrain inflation if supply expands. It can also increase inflation if households and businesses anticipate future gains and spend before supply catches up.

That is the AI version of the risk:

The boom can be disinflationary later and inflationary now.

Data centres need power now.

Chips need capital now.

Cloud capacity needs money now.

Specialist labour is scarce now.

Equity wealth effects happen now.

The productivity dividend arrives later, unevenly, and after measurement.

The AI productivity bet

AI can raise productivity in three ways.

First, it lets some people do more work:

A software engineer can test more options. A lawyer can review more documents. A finance team can reconcile more data. A founder can write faster. A customer-service team can respond sooner.

Second, it removes work that did not need to exist:

Summaries, formatting, first drafts, meeting notes, routine reporting and internal coordination consume large amounts of paid time. AI reduces some of that work quickly.

Third, it turns some human work into machine-executed output:

The “digital employee” does not take holiday, does not sleep, and can run across systems continuously.

Each channel has a cost:

The superworker can become more productive, then easier to copy.

The unnecessary task can disappear, then the salary attached to it disappears too.

The digital employee can look cheap, then the full bill arrives through tokens, inference, cloud, integration, supervision, verification and liability.

That is where many AI productivity claims are still weak:

They count the visible time saved.

They do not count the new cost system.

AI makes output cheaper. Verified output is harder. The more the machine produces, the more someone has to decide whether the work is correct, safe, legal, useful and worth acting on.

A company that produces five times more analysis and spends the same senior time reviewing it has not necessarily become five times more productive. It may have moved the bottleneck.

The K-shaped recovery

AI can raise GDP while making the lived economy feel worse.

That sentence will sound wrong to people who think in averages but it will sound familiar to people who live surrounded by bills.

The AI economy can split quickly:

Asset owners benefit when lower rates support equity prices, real estate and private company valuations. Large technology platforms benefit when compute, cloud, models and enterprise AI demand expand.

Highly skilled AI users benefit when they can multiply output and keep the gain.

Companies with clean workflows benefit before companies with fragmented data, unclear decision rights and weak management discipline.

Other groups experience a different economy:

Salary earners face weaker bargaining power when parts of their role become easier to automate.

Young workers face thinner entry-level pathways when firms use AI to remove junior work.

Middle managers face pressure when coordination work becomes visible and compressible.

Non-asset owners face higher prices without the asset appreciation that offsets them.

Local communities host data centres, power demand, water use and planning pressure while much of the economic gain accrues elsewhere.

That is the K-shaped AI economy.

This matters because legitimacy is built from the felt economy, not the aggregate chart.

A country can report productivity gains while large parts of the population feel poorer, less secure and less able to progress. That is a political problem, a workforce problem and a Board problem.

The inflationary redistribution

‘Shrinkflation’ is the consumer version of the problem.

The broader mechanism is inflationary redistribution:

If inflation stays above wage growth, households lose purchasing power.

If nominal growth rises, debt becomes easier to carry.

If interest rates are held below inflation for long enough, savers and fixed-income households lose real wealth.

If asset prices rise, owners of equities, homes and private businesses gain.

That is not a comfortable arithmetic:

The state gets debt relief.

Asset owners get valuation support.

Leveraged actors get easier real debt burdens.

Wage earners get higher prices.

Yet, nobody has to announce that policy. It can emerge from incentives:

High-debt governments prefer lower real rates.

Markets prefer cheaper money.

Borrowers prefer inflation that erodes debt.

Politicians prefer growth stories to austerity stories.

AI supplies the growth story.

Former Fed Chair Janet Yellen’s concern about Fed independence fits here. She warned that pressure on the Fed to hold down debt-service costs rather than focus on price stability can lead toward very high inflation. She also said Warsh understands the Fed, but enters the role under difficult political conditions.

The institutional risk is all too clear:

If AI productivity becomes the justification for lower rates before the productivity is visible, policy starts making a distributional bet.

The bet may pay off:

AI may lift output enough to reduce inflation pressure, raise real wages, improve margins and stabilise debt.

The bet may also fail:

If the gains arrive narrowly, the result is asset inflation, weaker labour bargaining power and a public told that technology has improved the economy while their own economy has deteriorated.

The productivity proof problem

AI productivity has to show up somewhere measurable and usage is not enough, prompts sent are not enough, documents generated are not enough, agents launched are not enough and headcount avoided is not enough if the old process remains and the review burden grows.

The useful metrics are harder:

• Output per hour.

• Unit cost per verified outcome.

• Revenue per employee.

• Error rates.

• Rework.

• Cycle time after review.

• Customer satisfaction after automation.

• Margin after full AI costs.

• Cost of control.

• Cost of failure.

• Cost of vendor dependency.

Let’s be clear:

A company has not proven AI productivity because employees use AI every week.

A company has proven something when the old work disappears, the new cost base is counted, quality holds, risk does not rise beyond appetite, and the gain reaches the P&L.

Governments face the same test.

AI has not solved a fiscal problem because speeches and forecasts assume higher productivity. It solves part of the fiscal problem only when measured output rises, tax receipts improve, public services become cheaper or better, and the gains are durable.

The current risk is that expectations are moving faster than proof which creates a gap:

Markets live in the expected gain.

Workers live in the present workflow.

Households live in the current price level.

Boards live somewhere between the above three.

What boards should ask now

Boards do not need an economic forecast to manage this risk.

They need discipline.

Ask:

1. Where has AI increased verified output, rather than produced more material?

2. Which productivity gains have reached the P&L?

3. Which AI costs were outside the original business case: compute, tokens, cloud, integration, review, legal, cyber, rework and vendor dependency?

4. Which workers capture the gains from AI: employees, management, shareholders, customers, vendors or the state?

5. Which roles are becoming more productive, and which are becoming easier to remove?

6. Which entry-level pathways are being weakened?

7. Which AI business cases depend on lower rates, high equity valuations or cheap compute staying available?

8. What happens if AI costs rise, rate cuts do not arrive, or regulation slows deployment?

9. Where does the company create local cost and remote benefit: power, water, land, labour displacement, customer distrust or political exposure?

10. What evidence would make the Board reduce its AI productivity assumptions?

The last question matters most. A productivity assumption with no failure test is a belief.

The management risk

The easiest mistake is to let the AI story become a ‘goal-seek’-style spreadsheet budget optimisation task:

A function wants lower cost growth.

AI will provide it.

A team wants lower headcount.

AI will provide it.

A strategy needs higher margins.

AI will provide them.

A debt forecast needs higher growth.

AI will provide it.

A rate-cut argument needs a non-inflationary supply shock.

AI will provide it.

The pattern is attractive because each claim may be partly true. That is how weak assumptions survive. They contain enough truth to avoid rejection.

A Board should make AI productivity compete with other capital-allocation claims. It should ask for evidence, timing, owners, costs and downside. It should ask whether the company is removing old processes or adding AI on top of them. It should ask whether the gain belongs to the company or to the vendor.

It should also ask what happens to trust.

Employees can use AI heavily while discounting its long-term value to them. They can improve output while fearing skill erosion. They can save time while seeing their own bargaining power weaken.

That is already part of the productivity equation.

Workforce trust is not a soft issue when the technology depends on the workforce to use, check, correct and integrate it.