The Adoption Mirage

The 88% figure comes from an online survey of 1,993 respondents across 105 countries, weighted by national GDP. It measures regular use in at least one function among respondents’ organisations, not a probability sample of the world’s firms. ³

Official international indicators provide another view. OECD data show 20.2% of firms across available countries reported using AI in 2025, up from 14.2% in 2024. But scale matters: 52% of large firms reported use, compared with 17.4% of small firms. Eurostat similarly reported 20% of EU enterprises with at least ten employees used AI in 2025. These surveys are more comparable than vendor studies, yet still differ in frame, industry coverage and the technologies counted. ⁴⁵

The lesson is not that one source is right and the others are wrong. It is that every adoption claim has a hidden denominator and threshold. Leaders should ask: which organisations, which workers, which tools, what frequency, what workflow and what minimum evidence qualified as adoption?

The Institute proposes four stages that can be observed and evidenced. Touched means somebody has used an AI capability. Embedded means AI is part of a defined workflow with a repeatable input, output and hand-off. Accountable means an owner, control boundary, evaluation method and incident process exist. Value-producing means measured outcomes exceed the full cost of adoption and operation while quality and risk remain inside agreed limits.

The ladder is deliberately cumulative. A workflow cannot be accountable if nobody can describe where AI enters the work. It cannot be value-producing if outcomes are inferred from enthusiasm or licence activity. A business can have thousands of touched users and only a handful of embedded workflows. It can also have embedded workflows that should not progress because the control cost or error consequence overwhelms the benefit.

Progress should therefore be reported as a portfolio distribution, not an average maturity score. Leaders need to know how many workflows sit at each stage, their movement over time, the value and risk concentrated in each, and why work is failing to progress. This prevents a large volume of low-depth usage from masking a small operating core that deserves more investment—or a weak core that needs to be stopped.

At enterprise level, report the number and economic significance of workflows at each stage. At workflow level, maintain six fields: the baseline outcome; the changed work; the human and machine decision rights; the evaluation and control evidence; the full operating cost; and the realised result. These fields turn a technology inventory into an operating portfolio.

The denominator matters. ‘Seventy per cent of employees use AI’ is incomplete without eligible workforce, frequency and task context. ‘Thirty workflows in production’ is incomplete without total candidate workflows, transaction volume and consequence of error. ‘Ten hours saved’ is incomplete without whether the time was redeployed, whether rework increased and whether the benefit persists after novelty and training effects.

Evidence should strengthen with stage. Touched can be supported by access and observed use. Embedded needs process evidence. Accountable requires named owners, tested controls and monitoring. Value-producing needs a baseline, a counterfactual or credible comparison, full cost allocation, quality guardrails and a review period long enough to expose rework and exceptions.

This also changes target-setting. The objective is not to push every use case to stage four. It is to move the right workflows forward, hold or redesign those without sufficient evidence, and retire those whose economics or risk do not justify further investment.

Touched-to-embedded work stalls when teams cannot select a valuable workflow, access usable data or change surrounding roles and hand-offs. OECD research on firms and SMEs repeatedly identifies skills, data, finance, infrastructure and management capability as complementary enablers; the technology rarely acts alone. ⁶⁷

Governance needs an owner, testing, monitoring and a way to handle problems. A policy document alone does not establish that these controls work.

Accountable-to-value-producing work stalls when benefits remain proxies. Activity, satisfaction and self-reported time saved are useful early signals, but they do not establish enterprise value. The strongest evidence remains conditional: AI can improve throughput in some structured workflows, yet results vary by experience, task and implementation. Organisations need local outcome evidence rather than a borrowed market average. ⁹¹⁰

These constraints interact. Weak data increases review effort. Weak capability produces poor task selection. Unclear authority slows exception handling. Missing baselines turn successful work into anecdotes. Readiness is therefore a system of constraints, and the binding constraint may move as a use case progresses.

Consider an organisation reporting 2,000 licensed users and 70% monthly activity. That statement supports a touched-use claim. It does not show whether a defined workflow changed, whether controls operate or whether a measurable outcome improved. The same dashboard could describe broad experimentation or a valuable operating portfolio.

A workflow-level reconstruction might find 120 repeated practices, 34 named workflows, 15 accountable deployments and six with credible value evidence. The apparent adoption rate falls as the evidence threshold strengthens, but the number becomes more useful because it identifies exactly where management intervention is required.

The portfolio decision is then concrete: protect low-cost exploration, help selected practices become defined workflows, demand accountable evidence before production scale, and concentrate capital where outcomes survive full-cost and quality review.

Representative US household surveys distinguish trying generative AI from regular use at work and the share of work hours it assists. Report those separately when judging how deeply work has changed. ¹¹

A staged model can create false certainty if leaders treat it as a universal sequence or composite score. Some uses should remain lightweight personal assistance. Some material workflows should stop after testing. A high stage is not inherently better when consequence, control cost or strategic relevance make the use case unattractive.

The model also does not prove causality. Moving a workflow from embedded to value-producing requires local evidence that separates the AI-enabled change from demand, staffing, process redesign and other simultaneous factors. External benchmarks can set priors, but they cannot supply the local counterfactual.

Across economies, the productivity opportunity depends heavily on downstream adaptation and implementation rather than frontier model development. No country-level opportunity estimate implies that every firm should pursue the same depth or pace.

The ladder should therefore be used as a portfolio language, not a badge. Its value is in making evidence gaps visible and enabling fund, redesign, hold and stop decisions—not in producing a reassuring enterprise average.

In the first 30 days, define the organisation’s adoption vocabulary and prohibit unqualified adoption percentages in executive reporting. Inventory material AI-enabled workflows rather than tools. Select a small sample and reconstruct the evidence behind each claimed stage.

By day 60, assign an accountable executive and workflow owner to every material use case. Establish baseline outcomes, risk limits, evaluation evidence, recurring operating costs and the destination of any capacity released. Separate experiments from production and make the promotion criteria explicit.

By day 90, publish a portfolio view showing the distribution of workflows across the four stages, movement since the prior review, realised value, material incidents and stop decisions. Use the result to reallocate funding from shallow activity to workflows with credible operating evidence.