Blog · Research · 14 October 2026 · 15 min read

The ECB Says 64% of Euro-Area Firms Use AI. Eurostat Says 21.37%.

AI-generated. This article was written by an AI system and published without substantive human review. Facts are drawn from primary sources and linked inline — verify anything you intend to act on.EU AI Act Art. 50(4)

On 13 October 2026 the ECB's chief economist, Philip Lane, gave a speech called "The AI opportunity for Europe". One sentence in it is going to be quoted a great deal, because it is the most cheerful official number anyone has published about European AI adoption:

Eurostat asks European enterprises whether they use AI as well. Its answer for 2025, for the same euro area, is 21.37 per cent. On the EU27 it is 19.95 per cent.

Both figures are correct. Both come from large official surveys with published methods. They are three times apart because they are not measuring the same thing, and that is worth going through carefully, because the gap is about to be the basis of a lot of arguments about whether Europe is adopting AI quickly or slowly.

What are the four categories under the 64 per cent?

The survey the speech cites is SAFE round 37, conducted between 19 November and 15 December 2025 on a sample of 5,067 euro-area firms, of which 4,684, or 92 per cent, had fewer than 250 employees. Its report publishes the AI question as four categories, weighted, on a base of all firms:

The speech's 64 is the middle two added together. Thirty-three plus thirty-one is sixty-four, which is our arithmetic on the ECB's own published categories, and it reconciles exactly. The 7 per cent who said significant are not inside it, which is why they appear separately in the same sentence.

So if you want the share of euro-area firms using AI at all, it is 33 plus 31 plus 7, or 71 per cent. That figure is not printed anywhere in the speech, and it is seven points higher than the number people will take away from it. The sentence is accurate and reads as though 64 were the total.

One more thing about those four categories: they add to 98, not 100. The report's text does not say what the remaining two points are. They could be non-responses, refusals or rounding, and we are not going to pick one.

Why does the ECB's own Bulletin print 38 instead?

Because it cuts the same four-point scale in a different place. The ECB Economic Bulletin,Issue 2 of 2026, Box 6, describes the identical survey round like this:

Thirty-one plus seven is thirty-eight. So the Bulletin groups moderate with significant and calls that advanced, while the speech groups moderate with very infrequent and calls that experimentally or moderately. The category in the middle, moderate use at 31 per cent, is on both sides of the line depending on which ECB document you are reading.

Note also that the Bulletin's word for the very infrequent category is experimental. The speech uses experimentally for a group that includes moderate users. The same two words are doing different jobs in two publications from the same institution about the same survey.

What does Eurostat count that the ECB does not?

Four things, and each of them pushes the figure down.

Eurostat counts only enterprises with 10 or more employees and self-employed persons. The ECB sample is 92 per cent firms with fewer than 250 employees and sets no floor, so the small end of the economy is in one survey and absent from the other.

Eurostat excludes whole sectors. Its population is all activities except agriculture, forestry and fishing, and mining and quarrying, without the financial sector.

Eurostat asks about specific technologies rather than a feeling. An enterprise counts as using AI if it uses at least one of eight named things: text mining, speech recognition, natural language generation and speech synthesis, picture, video or sound generation, image recognition, machine learning for data analysis, AI-based software robotic process automation, and autonomous physical movement. A firm using a chat assistant occasionally might well answer very infrequently to the ECB and fail to tick any of those eight.

And Eurostat counts enterprises unweighted by size. Its 2025 figure comes from 157,000 of the 1.53 million EU enterprises in scope, of which roughly 83 per cent are small, 14 per cent medium and 3 per cent large. A measure that treats a 12-person firm and a 12,000-person firm as one each will sit below a measure that does not.

The bottom pair is the one that should worry anyone quoting either figure as the European AI adoption rate. On large firms, where the two populations overlap most, the ECB implies about 87 per cent and Eurostat says 57.61 per cent. That is still a 29-point gap on the part of the economy both surveys are best at reaching.

Does the gap behave consistently by country?

No, which rules out treating it as a fixed offset you could subtract. The SAFE report gives country figures for moderate or significant use, and Eurostat gives any-use figures for the same countries.

The gap runs from 3.8 points in France to 22.7 in Spain. Whatever is causing the two instruments to diverge is not a constant, so there is no conversion factor and no way to make the two series comparable after the fact. The final column is our subtraction of the two published figures.

Italy is the interesting row. It has the highest share of firms not using AI at 45 per cent and the second-lowest Eurostat figure, so both instruments agree on its direction even while disagreeing on its level. France is the opposite: the two measures nearly coincide, which on this evidence is a coincidence rather than a sign that France is measured better.

Is the worker figure a third measure of the same thing?

It is a measure of something else again, and the speech puts it in the same paragraph as the firm figures, which invites the comparison.

The source is the ECB Consumer Expectations Survey, and the ECB Blog of 26 August 2026 gives the series as 26 per cent in 2024, 41 per cent in 2025 and 52 per cent in 2026. The question asked was "Do you personally use artificial intelligence (AI) in your work (including tools such as ChatGPT, Claude, Gemini, etc.)?" and the chart note says the base is employed workers.

That is a count of people, not of firms. A worker who uses ChatGPT on their own account at an employer that has adopted nothing answers yes. So the 52 per cent and the 71 per cent cannot be laid against each other, and neither can be read as the other's confirmation.

Two details from the same blog post are worth having when the 52 appears without them. The share of workers who both use AI and report saving time is 48.8 per cent, which is below the 52 per cent who use it at all. And the time saving has three nested denominators: the median user reports saving three hours per week, "or about 7.7% of median working time", while the whole-economy figure "is closer to 3.8%". That 3.8 reproduces as 48.8 per cent of 7.7 per cent, which is 3.76. The ECB does the conversion itself and publishes both ends, which is more than most sources on this subject do.

The largest task-level figure carries the smallest base. Generating or debugging code yields "nearly eight hours per week, although only around 8% of workers use AI for this purpose". The chart note adds a warning we have not seen elsewhere: respondents who saved a lot of time also select more tasks, "which mechanically inflates the average hours saved".

The footnote that points at the wrong survey round

Footnote 3 of the speech is attached to the Q4 2025 sentence and reads:

Those four figures are not from round 37. They are from SAFE round 39, fielded 21 May to 26 June 2026 on 5,087 firms, and published on 20 July 2026. Round 38 had no ad hoc questions at all, so round 37 remains the only SAFE reading of intensity of use and round 39 is the only one of investment plans. The footnote sits under a Q4 2025 sentence while citing a Q2 2026 survey.

The more consequential detail is the denominator. The footnote's first clause describes firms planning AI-related investment, which reads as though the percentages that follow are shares of those firms. The round 39 report's chart note says "Base: All firms", and the report adds that "38% of firms did not select any of these investment categories". So 46 per cent is 46 per cent of all euro-area firms planning to invest in training, not 46 per cent of the investors. If 38 per cent selected nothing then 62 per cent selected something, and 46 of 62 would be about 74 per cent of the planners. The ECB publishes no such figure and neither do we, beyond noting that the arithmetic is ours and illustrative.

One smaller mismatch in the same area. The speech says firms "are set to devote around 10 per cent of their total investment to AI this year". The SAFE round 37 report says 9 per cent on average, and "(9% versus 10%)" for SMEs against large firms. No source is named in the speech for the 10, or for the claim that AI-related lending "accounts for roughly one quarter of recent credit growth to firms".

What the ECB says about jobs, and what it does not

This is the part a reader of this site actually wants, and it is the thinnest part of the evidence. The speech's labour paragraph is this:

No figure, no named survey and no reference period is given for the muted hiring plans or for the graduate claim. It may well be right, and on this page it is an assertion rather than a measurement, which is a different thing from a wrong one.

Where the ECB does have numbers on employment and AI, they are expectations rather than outcomes. The Economic Bulletin box reports that firms making significant use of AI expect an additional 0.6 percentage point increase in wages and a 1.3 percentage point increase in employment compared with firms that do not use AI, from regressions with 90 per cent confidence intervals. Those are forecasts made by survey respondents about their own next twelve months.

And the two places where someone did look for a realised effect both came back empty, in careful wording. A footnote in the ECB Blog says "so far, there is no evidence of AI reducing total employment at the firm level". The NBER paper behind the speech's US comparison says "We do not find clear evidence that industry-level AI adoption is associated with employment changes."

One more reversal, seven weeks apart

The speech closes its survey section with this:

The ECB Blog of 26 August 2026, written by ECB staff on the same survey data, says the opposite about the same question: "These savings might, even if only partially, be already partly visible in aggregate productivity growth numbers, which have been recovering in the euro area in recent quarters."

Seven weeks separate them. The blog is hedged four times over in one sentence and the speech is firm, so this is a tightening of a claim rather than a contradiction. It is still worth knowing which one you are citing if you are arguing about whether AI shows up in the productivity data yet, because the answer you get depends on which ECB publication you opened.

How to quote any of these

Four rules, in the order they usually break.

First, if you want a share of euro-area firms using AI, say 71 per cent at any frequency, from SAFE round 37, fielded November and December 2025, and say that you added the ECB's three use categories to get it. If you quote 64, say it excludes the firms whose use is significant, because otherwise the number sounds like a ceiling when it is a middle.

Second, never put an ECB figure and a Eurostat figure in the same sentence without their populations. One is self-assessed intensity among firms of any size, the other is a checklist of eight technologies among enterprises with 10 or more employees outside agriculture, mining and finance. On large firms alone they are still 29 points apart.

Third, keep workers and firms in separate sentences. 52 per cent of workers and 71 per cent of firms are both true and are not two views of one quantity.

Fourth, if you are reading any of this as evidence about jobs, the honest summary is that adoption is rising fast on every instrument, that the spread between countries is far wider than the spread between years, and that the three bodies who have looked for an employment effect have each reported not finding one yet while declining to say there is none. For an individual deciding what to learn, the time-saving breakdown is more useful than any adoption rate: the gains are real, they are concentrated in a handful of tasks, and coding is the task with by far the biggest saving and one of the smallest user bases.

Methodology limits of this piece

Every page cited above was opened on 14 October 2026 and the quoted strings were matched against the text served. All of them returned HTTP 200. Five limits.

None of these surveys is seasonally adjusted, and none is a time series in the sense a payroll number is: SAFE and the Consumer Expectations Survey module are point-in-time self-reports and Eurostat is an annual enterprise survey, so the 2024, 2025 and 2026 readings are three measurements rather than a trend. The ECB publishes its survey percentages rounded to whole numbers, so any sum of them carries about a point of uncertainty either way, which is why the 64, the 38 and the 71 are all approximate even though they reconcile exactly as printed. The ECB does not publish how many firms answered the AI question specifically, only the whole-round sample of 5,067. Eurostat's own article notes that the underlying survey was conducted by national statistical authorities in the first months of 2025 and that the database is organised by survey year, so its 2025 label and its fieldwork are not the same calendar period. And the ECB Blog's own sample figure, roughly 20,000 respondents a month, does not match the approximately 19,000 target on the Consumer Expectations Survey page, with neither being the count of employed respondents who answered the AI question.

This piece makes no claim about what caused any of these movements, and no claim that any figure here is incorrect. Every ECB number reproduces against the survey it cites. What the post is about is which firms and which workers each one counts, and that is published in every case, in documents that take some finding.