One of the questions I hear most often in organisations is: How do we make people adopt AI?

It sounds practical, however I do think that is usually one question too late.

By the time adoption has become the objective, the organisation has already decided, often without noticing, that the technology deserves to spread before anyone has established which forms of work should change, which decisions may be delegated, what evidence would count as value or what new dependencies are being created.

Adoption describes diffusion, but it does not provide direction.

A company can distribute thousands of licences, train its entire workforce, record rapidly increasing prompt volumes and still preserve exactly the same workflows, approval structures, incentives and assumptions that existed before the technology arrived. The organisation may produce the same reports more quickly. It may also generate more presentations, more analyses, more meeting summaries and, in a triumph of technological progress, even more emails.

Does this necessarily means it has transformed? Absolutely not.

OpenAI’s 2026 B2B Signals report offers an instructive finding. Firms at the ninety-fifth percentile of usage generated approximately 3.5 times more tokens per worker than typical firms, up from twice as many one year earlier. Only 36% of the difference was explained by message volume. The remainder was associated with richer context, more complex work and greater use of advanced and agentic tools. OpenAI interprets generated tokens as a proxy for the amount of “intelligence demanded”, while explicitly acknowledging that they are not a direct measure of business value.

This distinction is more important than it first appears: frequency tells us that a tool is present in work, and depth suggests that some part of the work itself is migrating into the tool. However, neither tells us whether the migration was strategically desirable.

A company may consume more tokens because employees are solving more difficult problems. It may also consume more tokens because context is poorly designed, agents repeatedly fail, users regenerate unsatisfactory answers or the organisation has become remarkably efficient at producing industrial quantities of plausible prose. More AI is not yet better AI. It is simply more AI.

The same problem appears in OpenAI’s 2025 enterprise data, which reported an approximately 320-fold increase in reasoning-token consumption per organisation over twelve months. This is evidence that more computationally intensive models are entering real workflows. It is not evidence that organisational value increased by anything resembling the same factor.

The strategically relevant distinction is therefore not between organisations that use AI and those that do not.It is between organisations that add AI to inherited work and organisations that reconsider work because AI exists.The first adopts a technology. The second changes its theory of production.

Strategy is what separates them.

For most of modern corporate history, ‘intelligence’ entered the organisation through people: it arrived attached to salaries, professions, working hours, hierarchies, biographies and bodies that eventually required sleep.

Machine intelligence arrives differently. It can be purchased through an API, invoked inside a workflow and consumed in variable quantities. Its operational cost may depend on the context supplied, the model selected, the number of outputs generated, the internal reasoning performed, the tools called and the attempts required before a result is accepted.

The analogy between human and machine intelligence should not be stretched into metaphysics: a token is not a fragment of thought, and a model invocation is not an employee briefly coming into existence inside a data centre.

The change is economic: some forms of cognitive production have become metered.

A user sees a prompt and a response, but the underlying system may see input tokens, retrieved documents, cached context, internal reasoning, tool outputs, validation calls, retries and the histories required to preserve memory across interactions.

With an agent, a single instruction may become a tree of computational expenditure.

“Analyse this market” can mean searching sources, reading documents, generating hypotheses, writing code, inspecting results, revising assumptions and producing a report. What appears to the user as one task may be dozens of model-mediated operations.

This makes tokens economically significant… however, it does not make them economically sufficient.

Bergemann, Bonatti and Smolin model LLMs as technologies whose value depends on the allocation of input and output tokens across heterogeneous tasks. Tasks differ not only in their potential value, but also in their sensitivity to error. The implication is that there is no universally optimal level of token consumption. Efficient allocation depends on what the task is worth, how quality responds to additional computation and what happens when the system is wrong (Bergemann, Bonatti & Smolin, 2025).

Anthropic’s June 2026 Economic Index provides an empirical complement. In its usage data, conversations mapped to higher-wage occupations tended to consume more tokens and produce more computationally demanding artefacts. AI autonomy and token use also rose together across output categories. However, the relationship was noisy, and Anthropic correctly avoids treating token expenditure as a clean measure of economic value.

More valuable work may require more compute.

The inverse does not follow: more compute does not make work valuable.

Consolidated research on reasoning LLMs makes this asymmetry difficult to ignore: an inverted U-shaped relationship has been identified between chain-of-thought length and performance. That is, additional reasoning initially improves accuracy, but excessively long reasoning chains can introduce noise and reduce it (Wu et al., 2025).

An analogous pattern may emerge at organisational scale: more steps can create more opportunities for correction, however they can also create more opportunities for error, delay, duplication and unexamined delegation.

The strategic task is not to minimise token use. A system that spends twice as much but avoids a multimillion-euro mistake is not inefficient. Nor is the objective to maximise it, as though verbosity were a proxy for cognition and inference expenditure a corporate virtue.

The objective is to understand the marginal value of additional machine reasoning.When does more context improve the answer? When does another agent create useful specialisation? When does verification reduce risk? When does further reasoning merely produce a longer route to the same conclusion?

These are not questions for procurement alone, they are questions about the architecture of work.

Herbert Simon, just a ‘few’ years ago, observed that an abundance of information consumes the attention of those who receive it (Simon, 1971). I do believe that AI industrialises this problem.

It can produce alternatives, explanations, drafts, scenarios, images, code, recommendations and synthetic research at a speed no human organisation can meaningfully absorb. The immediate temptation is to describe this as abundance. The more precise description is a change in scarcity.

When writing becomes cheap, editing becomes scarce.When analysis becomes cheap, problem selection becomes scarce.When possible actions become cheap, judgement over consequences becomes scarce.When execution becomes increasingly autonomous, answerability becomes scarce.AI does not remove the bottleneck, rather it moves it downstream.

This matters because most adoption programmes are designed around the bottleneck that existed before AI, assuming that the main organisational problem is insufficient production: employees cannot write, analyse, code or search quickly enough.And clearly, in many workflows that is true.

But once generation accelerates, another constraint begins to dominate. Someone must determine which outputs deserve attention, which assumptions are valid, which exceptions matter, which recommendations may be trusted and which actions should occur. The cost of producing an answer falls, the cost of establishing that it is the right answer may not.

This is particularly visible in agentic systems. An agent does not merely generate content: each new capability may remove human effort from one layer while increasing the need for specification, monitoring, evaluation and exception management elsewhere. The interface becomes simpler, while the organisation behind it becomes more complex.

Hasan, Oettl and Samila describe this as a redistribution of complexity. General-purpose AI systems can appear remarkably simple to users while shifting complexity towards infrastructure, organisational controls, specialised expertise and the management of accuracy in high-stakes contexts (Hasan, Oettl & Samila, 2025).

Therefore, the complexity has not disappeared: it has become less visible to the person pressing Enter. This invisibility is strategically dangerous because hidden complexity tends to be mistaken for eliminated complexity: the organisation concludes that a task has become easy because the interface is easy.

However, in the last years we saw how a fluent interaction can, in reality, conceal poor retrieval, missing context, unstable evaluation, unmanaged permissions, ambiguous accountability and a growing tax of human review.

AI may save ten minutes of drafting and create twenty minutes of verification. It may accelerate one team and increase the coordination burden on three others, or produce a recommendation instantly while making it harder to reconstruct why the recommendation was accepted six months later.

These are not merely implementation defects; they are consequences of treating AI as an isolated tool rather than a new component in a socio-technical system. The relevant strategic question is therefore not: Where can AI generate something?It can generate something almost everywhere.The question is: Where does generation remove a genuine constraint without creating a more consequential one elsewhere?

Individual productivity is easier to observe than organisational transformation.A person writes an email in five minutes instead of twenty, a developer completes a function more quickly, a consultant drafts a report before lunch. The saved time feels real because it is real.

The mistake begins when this local improvement is treated as evidence that the entire production system has improved by the same amount. In a field experiment across 66 firms and 7,137 knowledge workers, Dillon et al. found that employees who actively used a generative AI tool spent approximately two fewer hours per week on email and reduced work outside normal hours. Yet the researchers did not detect meaningful changes in the quantity or composition of tasks performed as a result of individual access to the tool (Dillon et al., 2025).

AI changed part of the worker’s day… but it did not automatically change the organisation.

That is not a disappointing result: it is a clarification of the level at which value must be designed.

If a technology saves time inside an unchanged system, the organisation may simply refill the released capacity with more of the same work. Faster email produces more email. Faster slides produce more slides. Faster analysis produces a larger queue of analyses waiting for someone to decide whether any of them matter.Output rises. Value density may fall.

The jagged technological frontier identified by Dell’Acqua et al. reveals another complication. In their preregistered experiment involving 758 knowledge workers, participants using GPT-4 completed tasks within the model’s capability frontier more quickly and at higher quality. Yet on a complex task deliberately placed outside that frontier, AI users were 19% less likely to reach the correct answer (Dell’Acqua et al., 2026).

Hence, the same tool increased performance and reduced it.

The difference was not whether the employees had adopted AI: it was whether the task sat inside a capability boundary that users could not directly observe. This makes blanket adoption an intellectually weak strategy.

If performance varies by task, context and workflow position, then the organisation must develop the capacity to map where AI helps, where it misleads and how those boundaries move as models change. Otherwise, adoption amplifies both capability and error.

METR’s 2025 randomised study provides a useful warning against relying on perception alone. Experienced open-source developers working on familiar repositories took 19% longer when allowed to use then-current AI tools, despite expecting beforehand that AI would make them faster and continuing afterwards to believe that it had done so. The result should not be generalised to all software development, and METR itself later explained that broader adoption had made subsequent experiments vulnerable to selection effects. The durable insight is narrower: perceived productivity and measured productivity can diverge substantially (Becker et al., 2025; METR, 2026).

This divergence becomes even more difficult when AI changes which tasks people choose to perform.

Cunningham and Whitfill distinguish between uplift on old tasks, uplift on new tasks and uplift in value. A worker may become dramatically faster at a newly feasible task without creating a proportionate increase in economic value. When a task is performed mainly because AI has made it cheap, the researchers call it a “Cadillac Task” (Cunningham & Whitfill, 2026). A Cadillac, for those less invested in classic cars than I am, is a symbol of scale, luxury and conspicuous excess: impressive, desirable and far beyond what is strictly necessary.

Organisations will create many Cadillac tasks.

Reports that nobody previously requested.Analyses that were once too expensive and remain too irrelevant.Automations whose principal achievement is proving that they can be automated.

Human beings have never required assistance inventing unnecessary work. AI merely gives the practice infrastructure.The purpose of strategy is to prevent declining production costs from being mistaken for rising value.

The economic history of general-purpose technologies is rarely a story of immediate productivity.

Brynjolfsson, Rock and Syverson describe a productivity J-curve in which new technologies initially require substantial investments in intangible capital: redesigned processes, new products, organisational knowledge, skills and business models. The costs arrive before many of the benefits become measurable (Brynjolfsson, Rock & Syverson, 2021).

AI is no exception.

Giving employees access to a model is comparatively easy. Building an organisation capable of using it well is not.

The difficult work includes identifying which workflows matter, making relevant knowledge accessible, defining quality, constructing evaluations, redesigning roles, changing incentives, creating escalation mechanisms and learning from failures without converting every failure into a prohibition. These investments rarely appear in the price of the licence, but they are necessary in order to determine whether the licence becomes useful.

Anthropic’s March 2026 Economic Index offers evidence of learning-by-doing at the individual level. More experienced Claude users tended to attempt more complex and valuable work, collaborate more extensively with the model and obtain successful responses more often. Anthropic appropriately notes possible cohort and survivorship effects, but controlled analyses did not eliminate the observed relationship (Anthropic, 2026).

This matters because AI capability is not contained entirely inside the model: it emerges from an interaction between model capability, user capability and organisational context.

A powerful model with poor context may be less useful than a smaller model embedded in a well-designed workflow. As well as an agent with access to the wrong data may be more dangerous than a weaker system with a narrow and verified knowledge base.

The prompt is not the capability, the surrounding system is.

This is why AI literacy should not be reduced to teaching employees how to obtain better answers from a chatbot. The strategically important capability is learning how to recognise when AI changes the structure of a problem.

What information does the system require?Which part of the task is verifiable?Where does tacit knowledge enter?Which errors are reversible?Which decisions depend on values that the model cannot infer legitimately from historical data?What should happen to the time that automation releases?

The answers become institutional knowledge.Model access is rented, institutional learning accumulates.

That distinction is likely to matter more as foundation models become increasingly interchangeable for ordinary tasks. A competitor can purchase access to the same model within minutes… but it cannot instantly reproduce an organisation’s process knowledge, labelled exceptions, trusted evaluations, internal feedback loops, governance arrangements and understanding of where human judgement creates disproportionate value.

The strategic advantage is not possession of intelligence: I do believe It is the ability to compose intelligence with context, people and purpose.

The evolution from generative AI to agentic AI is often described as an increase in technical capability: few people note that is also a redistribution of decision rights.An assistant answers, an agent acts.

Between those two verbs lies most of the organisational problem.

Anthropic’s June 2026 data illustrates that autonomy is not a property of the underlying model alone. Across comparable output types, Claude Code sessions involved greater delegated autonomy than chat or Cowork interactions. The difference remained when Anthropic controlled for model class, suggesting that product architecture and mode of interaction can shape autonomy independently of raw model capability.This should be obvious: it is nevertheless routinely ignored.

Autonomy is produced by permissions, tools, memory, workflow design, stopping rules and the availability of human intervention. The same model may be a writing assistant in one interface and an operational actor in another.The relevant question is therefore not only:How intelligent is the model?It is:What has the surrounding system allowed that intelligence to do?

Economic work on human and artificial decision authority has already shown that allocation depends on comparative information, incentives and the ability to evaluate outcomes. More recent work by Gans distinguishes the value of strategic expertise from the value of coordinating interdependent choices, showing why influence, transparency and formal control become different mechanisms through which AI might shape organisational direction (Athey, Bryan & Gans, 2020; Gans, 2025).

This is where AI strategy becomes inseparable from governance.

Not governance understood as a final checklist applied after a system has been designed, but governance as the architecture of legitimate action.Who may propose?Who may execute?Who may interrupt?Who must explain?Who remains answerable when the action changes somebody else’s world?

“Human in the loop” is not an adequate answer: a human who receives hundreds of outputs without time, context or authority to challenge them is not exercising oversight. They are providing ceremonial approval to a decision system that has already become operationally sovereign.

Meaningful oversight requires the capacity to understand the basis of an output, identify uncertainty, contest assumptions, stop execution and accept responsibility for the result.The human must have more than proximity: the human must have agency.

This becomes especially important as AI systems produce increasingly persuasive explanations. Fluency can make a recommendation appear more inspectable than it really is. A generated rationale may describe a plausible path to the answer without constituting a faithful account of the internal process that produced it.The explanation can be generated, responsibility cannot.An organisation may delegate execution.It may delegate analysis.It may even delegate substantial portions of judgement.

But it cannot delegate away the fact that it designed the conditions under which the system was allowed to act.Responsibility does not disappear into the model: it returns through institutional design.

Inside a model there may be millions of representations. One may activate for the Golden Gate Bridge.But none is the exact configuration of what has happened to you, in the order in which it happened, with the loves, losses, humiliations, loyalties, travels, jokes, languages, and impossible decisions that made your mind yours.That configuration is not a performance.It is a life.Do not outsource the part of you that still has to become.

AI strategy is often represented as a roadmap of use cases: this is useful in approximately the same way that a list of possible journeys is useful without deciding where one intends to go.

A use case tells us that something can be done. Strategy explains why it deserves to exist, what it should change and what the organisation is prepared to become responsible for once it does.

The core strategic problem is allocation: an organisation now has access to multiple models, reasoning depths, retrieval systems, tools, agents and levels of autonomy. Each configuration carries a different combination of cost, latency, accuracy, explainability and risk.

This resembles capital allocation, except the resource being allocated is a probabilistic form of machine cognition.

Some tasks justify extensive reasoning because their value is high and their outputs can be verified.

Some require minimal compute because the answer is routine and additional reasoning contributes little.Some are inexpensive to automate but dangerous to execute without approval.Some should remain human not because machines are incapable of producing an answer, but because the act of deciding is itself part of the organisation’s responsibility.A mature strategy therefore connects at least three forms of economics.The economics of computation asks how much the system costs to run.The economics of production asks whether it improves the workflow.The economics of consequence asks what happens when the output enters the world.Cost per token belongs to the first.Cost per accepted outcome belongs to the second.Value per unit of reliable delegated intelligence belongs to the third.

The third is where these choices become strategic at the institutional level.

This is why model routing is not merely a technical optimisation. Choosing which model receives a task is an allocation of cognitive resources. Setting a reasoning budget is a decision about how much computational effort the problem deserves. Requiring verification is a decision about the cost of uncertainty. Preserving human approval is a decision about institutional authority.

Tokenomics becomes meaningful only when it is connected to these choices.

Otherwise, it is accounting for the machinery while remaining silent about what the machinery is producing.

The irony is that falling AI costs may make strategy more important, not less.

When the cost of generating an artefact approaches zero, organisations lose the economic friction that once forced them to choose carefully.More projects become feasible.More analyses can be commissioned.More software can be built.More decisions can be automated.The possibility space expands faster than the organisation’s ability to determine which possibilities matter.Cheap intelligence increases the cost of indiscrimination.The central strategic capability will therefore not be generating more options… machines will probably be excellent at that.It will be refusing options, sequencing them and binding them to a coherent direction.

Strategy is the practice through which abundance is made selective.

In these last months of thriving in AI Adoption, I understood that the organisations that thrive will not necessarily be those that deploy AI fastest. Nor will they be those that generate the most tokens, automate the most tasks or place an agent inside every process that management has not yet found time to understand.They will be the organisations that learn to distinguish cheap cognition from valuable judgement.They will recognise that AI does not simply reduce the cost of existing work: it changes which work becomes possible, which bottlenecks become decisive and which forms of authority can be embedded inside technical systems.Most importantly, they will understand that intelligence is not valuable in the abstract.It becomes valuable only when it is directed towards a purpose, constrained by reality and attached to responsibility.The machine can generate the analysis.It can propose the strategy.It can coordinate the tasks.It can even execute the decision.But the organisation still has to decide which future it is prepared to author through those actions.AI adoption asks whether an organisation is using intelligence, AI strategy asks whether the organisation has become intelligent about using it.When intelligence becomes a variable cost, strategy becomes the discipline of deciding what deserves to be thought, by whom, at what price and under whose responsibility.