Eurostat published last December a release that, on a different continent, would have been front-page news.
They were saying that 20% of European Union enterprises with at least ten employees now used artificial intelligence in some part of their business, up from 13.5 per cent the year before.
A jump of six and a half percentage points in twelve months. In Brussels, the number was greeted with quiet relief. In a Berlin think tank, an economist forwarded it to a colleague with a one-word comment: “finally.”
In a Bucharest co-working space, an SME owner read the same stats and did the maths on her own country. Romania came in at 5.2 per cent.
That spread, from Copenhagen at 42 per cent down to Bucharest at five, is where any honest editorial about European AI adoption has to begin.
The continent has not been standing still. It has been moving fast, in places, and not moving at all, in others. The aggregate twenty-per-cent number flatters and obscures in equal measure. It is the average of an economy that, on this question, no longer behaves like a single market.
The standard explanation for why Europe trails the United States on enterprise AI is regulatory. It is the AI Act, runs the line, that has spooked boards and tied up legal departments.
There is something to that, but not as much as the lobbyists would like. The deeper story is that European AI adoption is low for the same reasons European tech has been small for twenty years. Capital does not flow; skills are scarce.
The single market is single only on paper, and the firms that buy AI are still buying it, almost entirely, from American clouds.
Start with the capital. According to figures the OECD released in February and which Christine Lagarde cited in a speech to the European Parliament in November, roughly three-quarters of all AI venture capital in 2025 went to firms in the United States, totalling around $194 billion.
The European Union, taken together, attracted $15.8 billion. That is not a gap. That is two different orders of magnitude. The same speech leaned on Mario Draghi's earlier finding that around seventy per cent of the per-capita GDP gap between the EU and the United States is a productivity gap, and that the technology sector explains about two-thirds of that productivity gap since the turn of the century.
The numbers are not abstract. They are the reason a French SME thinking about an AI pilot reaches first for a budget that does not exist, and then for a service that does. The service is almost always American.
Which brings us to the second structural problem. Three US providers held roughly seventy per cent of the European cloud infrastructure market in 2025. European providers held about fifteen.
Every enterprise AI rollout in Europe that does not deliberately design around this fact ends up training on US compute, billed in dollars, governed by a foreign court's interpretation of data protection. This is not a hypothetical anxiety.
As we have documented, Mistral's chief executive Arthur Mensch has spent the past year arguing that Europe must “own and operate” its own AI infrastructure, and the company has put $830 million of debt behind a Paris data centre to make the point. Yet, it is also a long way from being delivered.
Inside firms, the limiting factor is people. The OECD's December 2025 report on AI adoption by small and medium-sized enterprises, prepared for the G7 presidency, found that half of all surveyed SMEs cite a skills shortage as their primary barrier to adoption. Forty per cent point to maintenance costs.
Thirty-two per cent flag hardware. Twenty-six per cent say they cannot understand the digital regulations they are meant to comply with. These are not the answers of executives who have been frightened out of AI by Brussels.
They are the answers of executives who would happily adopt AI tomorrow if they could find someone who could install it, run it and explain it in their own language. The Eurostat numbers reflect this. Large enterprises in the EU adopt AI at around fifty-five per cent.
Small ones sit at seventeen. The gap is not philosophical. It is the difference between having a data engineer in-house and not.
This is the point at which it becomes tempting, especially for an American reader, to cite the AI Act as proof that Europe has chosen process over progress. The honest reading is messier.
The Act's most invasive provisions, the ones covering high-risk systems, do not begin applying until August 2026. The European Commission has already moved to soften the edges: in a Digital Omnibus proposal published on 19 November 2025, it set a target of reducing compliance burden by twenty-five per cent overall and thirty-five per cent for SMEs by 2029, and extended the simplified SME framework to firms with up to 750 employees and €150 million in turnover.
The Commission has clearly read the same survey data. Whether it has read it in time is another question. Industry analyses suggest EU and UK developers report launch delays in nearly six in ten cases because of the Act, and that something approaching two-thirds of European companies still cannot articulate what their obligations under it are.
Regulation is not the main thing slowing European AI adoption. But it is not nothing, and pretending it is would be a different kind of dishonesty.
Set against this, the bright spots are real and underreported. Denmark's enterprise AI adoption is now higher than the US enterprise average reported by Stanford. Finland and Sweden are not far behind.
McKinsey's State of AI 2025 survey, with nearly two thousand respondents across 105 countries, found that 88 per cent of organisations globally now regularly use AI in at least one function. O
nly six per cent, however, are seeing material enterprise-wide impact, defined as a five-per-cent or greater contribution to EBIT. On that second measure, the European laggard problem is less severe than the headline numbers suggest. The Americans are running pilots too.
They are simply running more of them. What separates the high performers everywhere is not country but commitment: senior-leadership ownership, end-to-end workflow redesign, and a willingness to spend money on infrastructure before measuring returns.
Those are habits, not regulations. Europe can choose them at any time.
It is also worth saying that European industry is not absent from the productive end of the curve. Siemens has spent two years pushing its Industrial Copilot into factory-floor workflows, with new agentic capabilities announced at Automate 2025.
SAP has woven Joule into its core ERP. Mistral has signed multi-year deployment deals with Accenture and at least one major European bank. The picture is not one of paralysis. It is one of unevenness, and the unevenness has a shape.
The firms doing AI well in Europe are large, well-capitalised, internationally minded, and concentrated in a handful of countries.
The firms not doing AI are small, regionally bound, and disproportionately in the East and South. The single market, on this technology, is two markets.
If there is a real bottleneck, that is it. Not Brussels, not chip shortages, not Mark Zuckerberg's purchasing power.
I think it is the absence of a European capital and skills base that allows a Slovenian logistics firm or a Portuguese clinic to adopt AI as easily as a Danish bank already does.
The AI Act will get its share of blame and some of it will be earned. But the more durable failure is older and has nothing to do with AI. It is the failure to finish the single market for capital, for skills and for cloud infrastructure that Mario Draghi spent four hundred pages describing last year, and that successive European Councils have responded to with communiqués and pilot programmes.
Probably, Eurostat will publish another in a year, and another in two. If the gap between Denmark and Romania narrows, it will be because Europe finally decided that adopting AI was a question about industrial policy and human capital rather than a question about ethics frameworks.
If the gap widens, the explanation will be sitting in the same survey it has been sitting in for a decade.