Enterprise AI adoption

01 · The problem with the evidence

Four numbers, one word

If you sit on a board in Europe, you have been shown at least two of these in the last six months, usually on the same slide, usually without their denominators.

Every one of those figures is defensible on its own terms. Put them on the same axis and they are incoherent. Adoption cannot simultaneously be 88% and 20%. Failure cannot be 95% while half of all American businesses are voluntarily paying for the thing every month and renewing.

The reflex is to assume someone is lying, or that the gap is Europe versus America. Neither is right. The four numbers disagree because they count four different populations doing four different things. One asks whether AI is used anywhere; one asks whether a pilot reached structural production; one records whether an invoice was paid; one samples the actual population of European firms. Each uses a different verb, and the hardest of them — production — is never defined in a way another study could reuse.

02 · The definition problem

Nobody defines "production"

Here is the same four figures with their actual measurement basis attached. This table is the single most useful thing in this article.

Table scrolls sideways →

Read down the "what it actually counted" column and the contradiction largely dissolves — so here is the arithmetic rather than the assertion. McKinsey's 88% and Eurostat's 20% are not rival estimates of one quantity. McKinsey surveys a self-selected panel of largely large, largely digitised organisations and asks whether AI is used anywhere. Eurostat samples the whole population of European firms with ten or more staff, which is overwhelmingly small companies.

Put them on the same footing and the gap shrinks to something explicable. Restrict Eurostat to enterprises with 250 or more employees — roughly the population McKinsey is sampling from — and European adoption is 55.0%, not 20%. The remaining distance to 88% is the combination of a global rather than European frame, a self-selected respondent panel, and Europe genuinely trailing on this measure. That is a difference of degree between comparable things, not a contradiction. The 20-versus-88 version of the argument is an artefact of comparing a population statistic with a panel statistic.

The deeper problem is the missing definition. Across the sources consulted for this piece — listed in full at the foot — not one defines production in a way another source could reuse. Is a workload in production when it serves external customers? When it has an on-call rotation? When it has a budget line? When finance can attribute a cost to it? Each study picks an implicit answer, none states it, and the resulting rates are then compared as though they were commensurable.

03 · The most-quoted number

The 95% figure, examined

One statistic dominates this topic. It appears in board papers, vendor decks and the 2026 commentary of firms that have their own contradictory numbers to sell — we compared those side by side in the big consulting AI frameworks, benchmarked.

The claim — that roughly 95% of enterprise GenAI pilots produce no measurable P&L return — originates in The GenAI Divide: State of AI in Business 2025, published by MIT's Project NANDA in July–August 2025. Its evidence base is 52 structured interviews, a survey of 153 leaders, and a review of around 300 public deployments.

• The sample cannot carry the headline. Fifty-two interviews over a six-month window is a qualitative study. It supports a hypothesis about mechanism. It does not support a population percentage quoted to two significant figures.
• "Failure" is defined broadly. The report counts anything short of structural production deployment as no-return, which sweeps in pilots that concluded correctly — a well-run pilot that answers "no, not worth scaling" is a success of governance, not a failure of AI.
• It is not peer-reviewed, and the publisher has a commercial position. NANDA operates paid corporate memberships. That does not make the finding wrong. It does mean the framing was not produced by a disinterested party, which is exactly the standard the report applies to vendors.
• The primary source is now hard to reach. As of this article's snapshot date, the official MIT NANDA URL no longer serves the report directly. The document circulates as a mirrored PDF. A statistic this influential should be easier to check than it is.

And yet the report's central mechanism is the most interesting thing anyone published in 2025, and nobody quotes it. NANDA's argument concerns workflow rather than model quality: the failure driver is tools that do not retain feedback or adapt — and, relatedly, that shadow AI use runs at over 90% of firms while official LLM subscriptions sit near 40%. Employees are getting value from AI that their employer's programme is not capturing.

The same discipline applies to two Gartner figures in wide circulation. That over 40% of agentic AI projects will be cancelled by end-2027, and that at least 30% of GenAI projects would be abandoned after proof-of-concept by end-2025, are both predictions — published June 2025 and July 2024 respectively. They are routinely quoted as though they were observed outcomes.

The end-2025 one has now passed its own test date, so we went looking for the scoring. We could not find one. Gartner does not appear to have published any public assessment of whether its 30% prediction held, and the figure most often produced as confirmation — S&P Global Market Intelligence's "42% of companies abandoned most of their AI initiatives in 2025" — we could only reach through secondary summaries; the primary page did not resolve at our snapshot date, so we do not cite it as fact either.

04 · The evidence that holds

What survives scrutiny

Strip out the self-reported, the predicted and the unverifiable, and a smaller but sturdier picture remains.

The Ramp figure deserves more attention than it gets. It is derived from corporate card and spend ledger data rather than from asking executives what they think. Someone in the business made a purchasing decision and finance settled the invoice. It crossed 50% for the first time in March 2026, up from around 35% a year earlier. It is narrow — US only, Ramp's own customer base, and a paid subscription proves purchase rather than value — but no one self-reported it, which makes it unique here.

Of the failure literature, RAND's 2024 study is the most methodologically honest. Built from 65 semi-structured interviews with data scientists and engineers across government and industry, it puts AI project failure above 80%, roughly twice the rate of non-AI IT projects. Crucially, 84% of industry interviewees attributed the primary cause to leadership-driven issues rather than technology — which is a strategy and leadership problem wearing a technology costume. It is still a qualitative study and should be read as one, but it names its sample, states its method, and its conclusion points at the thing executives control.

Then there is the number that should stop a board mid-sentence. From BCG's survey of 1,803 C-level executives across 19 markets: 60% of companies define and monitor no financial KPI at all for AI value. Building that instrument is its own discipline, and the one we run as KPI setup. Before any argument about whether AI pays, most organisations have not built the instrument that would tell them.

Every figure in this article, graded

Size is the wrong axis. What produced a number decides what it can support. Official statistics and transaction records are evidence. Executive surveys are perception. Predictions are neither.

Table scrolls sideways →

Printed and held next to the next AI deck you are shown, that table is most of the value of this article.

05 · The European picture

What Europe's own numbers say

Almost every statistic in the global AI adoption debate comes from a US-centric vendor or consultancy survey. Europe has something better and it is barely cited: an official statistical office that measures the actual population rather than a panel.

Eurostat's 2025 figures, published 11 December 2025, cover enterprises with ten or more employees across all 27 member states. And they carry the size breakdown that almost every citation of them drops.

Stay on this one, because it is the mistake this article was written to warn against and Europe's own data makes it easy to commit. If you run a company with more than 250 employees, your peer adoption rate is 55%, not 20%. Roughly half your competitors are already using AI somewhere in the business. The widely-quoted European figure describes an economy dominated by small firms, and reading it as your benchmark will tell you that you are ahead when you are behind.

Two further things the Eurostat data says, both of which cut against the vendor narrative.

• The country spread is enormous. Denmark 42.0%, Finland 37.8% and Sweden 35.0% at the top; Romania 5.2%, Poland 8.4% and Bulgaria 8.5% at the bottom. An eightfold gap between member states means "European AI adoption" as a single number is close to meaningless for planning. We work out of Warsaw, so Poland sitting second-from-bottom is a fact we plan around rather than one we argue with; the wider context is in our 2026 Polish economy briefing.
• The barriers are not the ones the vendor literature emphasises. See the chart below.

That last line matters for anyone buying AI services in Europe. The dominant blocker is not scepticism about value — only one firm in five said AI would not be useful. It is that companies do not have the people who would know how to do it, by nearly twenty points over the next reason. Capability gaps are not closed by buying a platform, which is the whole argument for senior interim cover over a licence and a training budget.

06 · Causes, not symptoms

Why adoption stalls

The literature converges here more than anywhere else.

McKinsey's November 2025 barrier data puts regulatory and legal concerns at 44–48%, AI risk including bias and IP at 44–47%, change management and organisational silos at 39–40%, legacy technology at 30–33%, and strategy gaps at 28–32%. Separately, 51% of respondents reported at least one negative AI outcome, with inaccuracy the most common at 30%.

RAND's interviews land on leadership. Eurostat's population data lands on expertise. BCG's finding lands on the absence of a financial KPI. Read together, the picture is consistent and unflattering: the constraint is organisational capability and executive attention, not model quality.

This is the point at which most consulting content produces a maturity model and invites you to locate yourself on it. We are not going to do that, and the reason is blunt: a maturity model tells you where you sit relative to a curve someone else drew, and generates anxiety rather than a decision. The three things the evidence actually supports are narrower and duller.

07 · What is genuinely new

What actually changed in 2026

Three things changed in a way that should alter a plan written in 2025. Several others are being marketed as change and are not.

Inference cost collapsed, and it is documented

The Stanford HAI AI Index, published 13 April 2026, reports that LLM inference prices fell at a median rate of roughly 50-fold per year across capability tiers, accelerating to around 200-fold per year for some tiers since January 2024. This is the best-sourced cost claim in the field. It does not mean bills are falling — consumption has grown faster than unit price has dropped, which is why cost measurement matters more rather than less.

Agents moved from demo to deployment, unevenly

McKinsey's November 2025 survey has 62% of organisations experimenting with AI agents and 23% scaling them in at least one function. That is a real shift from 2024. It is also the area where the marketing outruns the evidence most severely — the widely-circulated claim that "88% of agent pilots fail to graduate to production" traces only to SEO aggregator sites with no underlying study, and we do not cite it. The build-versus-buy decision underneath agent platforms is covered properly in our agent platform playbook.

The European regulatory clock moved

The Digital Omnibus reset the high-risk timetable, as above. The practical effect is a reallocation rather than a reprieve: classification and conformity work for Annex III systems has roughly two more years of runway than most 2025 plans assumed, while GPAI obligations have been live since August 2025 and are the nearer exposure for anyone building on a foundation model. Programmes scoped in 2025 frequently have these the wrong way round.

08 · Practical

How to read this as the buyer

If you are a CEO, CIO, board member or PE operating partner, the useful output of all this is a short list of questions that cut through a pitch faster than an RFP.

Table scrolls sideways →

And two more things about the state of the field, both gaps rather than findings:

• We could not find an audited study. Across the sources below we found no general-ledger-verified, cross-industry EBITDA attribution of AI spend. If one exists we would like to see it; until then, every "value realised" figure we located is executive perception, however large the sample.
• The rates are not comparable. Because none of these sources defines production consistently, their adoption and failure percentages cannot be placed on a common axis — which has not stopped anyone from doing exactly that.

09 · Us

Where Consulting Huber fits

We are two senior operators. We do not sell an AI platform and have no licence revenue riding on the answer, which is the only reason to trust a word of the above — we had nothing to gain from any particular number winning.

We also cap engagements at four a quarter. That is a policy, not a result, and you should read it as one: it is how the people who scoped the work stay the people who do it. Every engagement carries a named duration and a written exit criterion, which is a commercial commitment and a methodological one at once — work that cannot say when it ends cannot say what it was for. What we have actually delivered, with the numbers attached and the confidential parts marked as confidential, is in four engagements, real numbers. Judge that rather than this paragraph.

If you want to go deeper on a specific part of this:

• The AI measurement crisis — what enterprise AI actually costs and how to instrument cost-per-outcome.
• The AI value creation playbook — the four value levers and the 90-day pattern from pilot to measured impact.
• The delivery layer underneath AI — the organisational prerequisites, by worked example.
• Enterprise AI agents: build, buy or open source — the platform decision, with an exit-cost scorer.
• The EU AI Act operator's playbook — risk tiers, the revised deadlines and a 90-day plan.