AI adoption is earlier than it looks—and later than it looks

The taxi-driver signal is not really about timing the market

The old market saying holds that by the time a taxi driver is talking about shares, the bull market is close to its end. It is a poor trading rule, but a useful warning about social normalisation. Once an opportunity is being discussed everywhere, mere awareness of it is no longer a source of advantage.

An advert for a course promising to keep managers relevant in an AI-enabled organisation can produce the same reaction. If AI literacy is now being sold as protection against professional irrelevance, perhaps anyone arriving only for that reason is already late.

That instinct is directionally sound, but incomplete. Managers may be late to gain distinction merely from recognising that AI matters. They are not necessarily late to build the capability required to use it well. In many organisations, the difficult work of redesigning processes, allocating authority, testing evidence, managing exceptions and capturing value has barely begun.

AI can therefore be late as a subject that leaders are expected to understand and early as an operating capability. The apparent contradiction comes from using one word, adoption, to describe several different transitions.

Headline adoption figures conceal the depth problem

AI adoption can look advanced or embryonic depending on what the survey counts.

The Office for National Statistics reported that around 35 per cent of UK businesses with 10 or more employees used at least one AI technology in June 2026, up from around 12 per cent in late 2023. That is rapid expansion. The same analysis found that the average number of AI technologies used by adopting businesses had moved only modestly, from around 1.4 to 1.6. Only 10 per cent of adopters described their AI use as extensive, and only 15 per cent said more than half of employees used AI as part of their daily work.¹

The United States shows a similar pattern. Census Bureau data for late 2025 and early 2026 found that 18 per cent of firms used AI in a business function. Among the adopters, 57 per cent used it in no more than three business functions and 65 per cent limited use to no more than three worker tasks. Writing, document analysis and information search led the list.²

Across the European Union, 20 per cent of enterprises with at least 10 employees reported using AI in 2025. The figure rose to 55 per cent among large enterprises, but remained 17 per cent among small ones.³

None of those figures says that AI is trivial. They say that breadth and depth are different. A business can count as an adopter because some employees use a language model, one function has bought an AI-enabled application, or a small pilot exists. That is contact with the technology. It is not necessarily a redesigned operating model.

This is why very high executive-survey estimates and much lower official business-survey estimates can both be honest. A senior respondent may know that the organisation uses AI somewhere. A worker-level or function-level measure may reveal that the use is narrow, informal or lightly embedded. Adoption rates often measure the first touch, while leaders care about the distance travelled afterwards.

The institutional machinery is arriving faster than the operating capability

While operational depth remains limited, the organisational machinery around AI is already multiplying.

The International Association of Privacy Professionals reported in 2025 that 77 per cent of surveyed organisations were working on AI governance. Among organisations already using AI, the figure approached 90 per cent. More strikingly, 30 per cent of organisations not yet using AI said they were already working on governance.⁴

That survey should not be read as a population estimate. It drew responses from more than 670 people in 45 countries and territories, many of them working in privacy, compliance and related governance fields. The sample is naturally more governance-aware than the average organisation. But that makes one part of the finding especially useful: it documents a real governance-first pattern in which structures are created before AI becomes a strategic or operational capability.

Training shows the same uneven progression. ONS data indicate that training or retraining existing staff is the most common way UK businesses integrate AI skills. Around 40 per cent of medium-sized and large businesses reported doing so. Yet only 11 per cent of businesses with 10 or more employees said more than half of their workforce had received AI-related training.¹

The wider institutional environment is also moving. The EU AI Act requires providers and deployers to take measures supporting staff AI literacy, with additional training expectations around human oversight of high-risk systems. The rules do not require a particular officer, committee or certificate, but they make AI literacy an organisational responsibility rather than a personal hobby.⁵

ISO/IEC 42001 now provides an international management-system framework for defining AI responsibilities, policies, risk controls, monitoring and continual improvement.⁶ US federal agencies have been directed to designate Chief AI Officers and distribute AI responsibilities through governance and acquisition structures.⁷

Roles, policies, councils, standards, training programmes, inventories and procurement rules can therefore appear before widespread workflow transformation. The scaffolding may be visible while much of the building is still missing.

Three curves describe adoption better than one

The simplest way to resolve the paradox is to separate three forms of progress.

1. Awareness

Awareness means that people recognise AI as relevant. Boards discuss it. Employees use it. Vendors place it in every proposal. Managers worry about being left behind. Courses, conferences and qualifications appear. Regulators and professional bodies publish guidance.

Awareness can spread extremely quickly because it requires little organisational change. A leader can become aware in an afternoon. An employee can begin experimenting within minutes. A market can become saturated with AI language long before most organisations have changed a core process.

Awareness is already far enough advanced that generic AI literacy is losing its power as a differentiator. Knowing that AI matters is becoming table stakes.

2. Institutionalisation

Institutionalisation begins when the organisation gives AI a place in its formal machinery. It names an owner, sets a policy, creates a council, allocates a budget, selects approved tools, adds review gates, commissions training, maps regulation or adopts a standard.

These steps can be prudent and necessary. They can also be produced faster than operational competence. A policy can be written before exceptions are understood. A committee can be formed before it has useful evidence to consider. A training course can be purchased before managers know which workflows need redesigning.

Institutionalisation is evidence that the organisation expects AI to matter. It is not proof that the organisation knows how to make it matter.

3. Operationalisation

Operationalisation is the harder curve. AI becomes part of repeatable work rather than an isolated interaction. The organisation knows what the system is meant to do, where its inputs come from, who checks its outputs, how exceptions are handled, what happens when the model or supplier changes, how performance is measured and who remains accountable for the result.

This is where adoption moves from use to capability.

It usually requires process knowledge, data access, integration, evaluation, workflow redesign, employee confidence, local judgement, management attention and a credible route from experiment to supported operation. Those elements cannot be created by announcing a policy or attending a course.

The third curve is slower because it changes the organisation rather than its vocabulary.

Why the scaffolding can run ahead

There are good reasons for institutionalisation to precede operationalisation.

First, AI arrives with visible risk. Leaders do not need a mature deployment to anticipate problems involving confidential data, inaccurate output, bias, intellectual property, supplier dependence or unclear accountability. Governance can rationally begin before scale.

Second, adoption often starts with employees rather than strategy. The ONS found that reported individual use was higher than business-level use, consistent with informal or ad hoc activity inside organisations.¹ Policies and training may therefore be responses to use that already exists but is not yet recognised as a formal workflow.

Third, organisational structures are comparatively cheap. Creating an AI council, buying a training package or adding a review question to procurement is easier than rebuilding a service process, integrating systems, measuring outcomes and changing who decides what.

Fourth, an ecosystem has formed around preparation. Regulators, standards bodies, consultancies, technology vendors, professional associations, universities and training providers can package guidance and credentials faster than individual organisations can discover how AI changes their particular work.

Finally, visible structures reassure senior leaders that the issue is being addressed. That reassurance may be justified. It may also allow activity to be mistaken for capability.

The scaffolding gap is not automatically a failure

A gap between institutionalisation and operationalisation is not necessarily evidence of waste or theatre. Scaffolding is supposed to arrive before the building is complete.

Clear ownership, minimum controls, approved tools and targeted training can make experimentation safer and reduce duplicated effort. A governance process can help a useful pilot survive procurement, security and legal review. A central team can create shared evaluation methods that individual functions could not develop alone.

The problem arises when the temporary support structure becomes the output.

An organisation can count policies, training completions, steering meetings, licences and pilots while remaining unable to move a use case into dependable operation. It can appoint a Chief AI Officer whose office becomes the destination for every unresolved decision. It can require approval so cumbersome that employees simply continue using unapproved tools. It can teach prompt techniques without changing a single workflow or management expectation.

The volume of structure then creates false confidence. Leaders see evidence of motion and assume capability is advancing at the same rate.

This is why the quality question matters later, even if it is not needed to establish the volume point. A structure is only useful when it helps the organisation make better decisions and build better work.

The Three-Curve AI Adoption Test

Use this test to distinguish visible preparation from operational capability. The aim is not to score the organisation with false precision. It is to identify which curve is ahead and what management problem follows.

This is an editorial diagnostic, not a scored maturity model. Read the pattern of evidence across the three curves.

Read the pattern, not just the score

If awareness is ahead of institutionalisation, the organisation is likely to have shadow AI, inconsistent practices and unmanaged risk.

If institutionalisation is ahead of operationalisation, the organisation has a scaffolding gap. The next task is not another general policy or awareness course. It is to create a route from promising use case to supported workflow.

If operationalisation is ahead of institutionalisation, useful work may be happening without sufficient visibility, control, shared learning or resilience. Success itself can become a hidden dependency.

If the three curves are moving together, the organisation is more likely to scale deliberately: awareness creates demand, institutionalisation creates coherence, and operationalisation produces evidence that improves the next round of decisions.

What this means for the manager who feels late

The manager who begins an AI course today is not too late to learn. They may be too late to gain much distinction simply from learning that AI exists, using a chatbot or acquiring generic vocabulary.

The differentiator is moving.

It once sat in access: who had tried the tools. It then moved to literacy: who understood their possibilities and limitations. It is now moving towards operational judgement: who can identify a suitable workflow, define the outcome, design the handovers, preserve accountability, test quality, manage exceptions and show whether the new process is better than the old one.

That is not a specialist technical skill alone. It is management applied to a new class of capability.

The same logic explains why the Chief AI Officer can be both necessary and transitional. The role is useful while the institutional structure has to be connected to real work. If it succeeds, more of that responsibility should become ordinary management rather than remain trapped in a specialist office.

AI literacy is becoming expected before AI operational competence has become common. That is why the manager can be late to the conversation and early to the difficult work at the same time.

Stop asking whether adoption is early or late

The question is too blunt.

AI awareness is already mainstream enough to shape careers, training markets, board agendas and regulation. Institutionalisation is advanced enough to produce roles, policies, councils, standards and governance programmes in considerable volume. Operationalisation remains patchy, narrow and difficult to measure.

Those three facts can coexist.

The practical question for a leader is not, "Are we early or late?" It is, "Which curve are we looking at, which one is lagging, and what would bring them back into alignment?"

The taxi-driver proverb is useful only up to a point. Widespread conversation tells us that the subject has become ordinary. It does not tell us that organisations have changed how they work.

AI adoption is earlier than the surrounding machinery makes it appear, and later than the operational evidence suggests. Leaders who can distinguish awareness, institutionalisation and operationalisation will be less likely to confuse activity with capability, or caution with delay.