One in five EU firms now uses AI. Company size sets the pace.
Eurostat’s 2025 survey of 157,000 enterprises records a rise of 6.5 percentage points in twelve months. The gradient by company size barely shifted, and the obstacle smaller firms report is a lack of expertise rather than doubt about the technology.
Gli Insights sono pubblicati in inglese.
Eurostat published the 2025 results of the EU survey on ICT usage and e-commerce in enterprises on 11 December 2025. The headline is a rate of change: 20.0% of EU enterprises with ten or more persons employed used artificial intelligence technologies in 2025, against 13.5% a year earlier — a rise of 6.5 percentage points, measured across 157,000 enterprises surveyed out of the 1.53 million within the survey’s scope.
The aggregate moves as a single number. The population beneath it does not. Among large enterprises with 250 or more staff, 55.03% used AI technologies. Among medium-sized enterprises, 30.36%. Among small enterprises with ten to forty-nine employees, 17%. Two companies selling comparable products to comparable customers are separated, in these statistics, by a factor of more than three on headcount alone.
That distance is the subject worth examining, because the same survey records why firms that stayed out did so — and the reason they give is not the one usually assumed.
Adoption is rising fastest where it was already highest
National rates range from 42.03% in Denmark, 37.82% in Finland and 35.04% in Sweden down to 8.55% in Bulgaria, 8.36% in Poland and 5.21% in Romania. The largest annual increases were recorded in Denmark, up 14.5 percentage points, Finland, up 13.5, and Lithuania, up 12.5 — three countries that already stood well above the European average. The sectoral picture is equally concentrated: 62.52% of enterprises in information and communication used AI technologies, against 40.43% in professional, scientific and technical activities.
What firms do with it is narrower than the term suggests. The four most widely used technologies in 2025 were the analysis of written language, reported by 11.75% of enterprises, the generation of images, video or audio (9.55%), the generation of written or spoken language (8.76%) and the conversion of spoken language into machine-readable form (7.2%). Among enterprises that use AI at all, the leading applications are marketing and sales, cited by 34.70%, and business administration processes, cited by 31.05%. European adoption, in its present phase, is largely the application of language tools to commercial and administrative text.
The barrier smaller firms report is expertise, not scepticism
Eurostat asks enterprises that considered using AI but did not adopt it what held them back. Among that group, 70.89% cited a lack of relevant expertise, 52.52% a lack of clarity about the legal consequences and 48.83% concerns about data protection and privacy. Only 20.68% concluded that AI was not useful for their enterprise.
That last figure deserves attention. Roughly four in five firms that examined the question did not reject the technology on its merits; they stopped at their own capacity to carry it. A constraint of that kind behaves differently from disbelief. It does not dissolve as tools become cheaper or more capable, and it is not addressed by another round of demonstrations.
Access to the infrastructure has diffused; the ability to use it has not
The comparison with cloud computing makes the point precisely. In 2025, 49.3% of small EU enterprises bought paid cloud services, against 66.78% of medium-sized and 84.67% of large ones, on an EU average of 52.74%. Set that against AI use of 17% and 55.03%. On cloud, small firms stand at roughly 58% of the large-firm rate; on AI, at 31%. Where the requirement is a subscription, company size matters comparatively little. Where it is judgement about how work should be reorganised, it matters a great deal.
The staffing statistics point the same way. In 2024, 20.05% of EU enterprises employed ICT specialists: 14.04% of small enterprises against 78.44% of large ones. Training follows the same slope. In 2023, 7.11% of small enterprises provided training to develop the ICT skills of their ICT specialists, against 61.41% of large enterprises and an EU average of 11.44%. The labour market offers no easy correction either: in the same year 6.23% of small enterprises recruited or tried to recruit ICT specialists, against 51.87% of large ones, and 57.5% of the enterprises that tried reported difficulty in filling the vacancies.
Switzerland shows the same gradient from a separate statistical system
Switzerland sits outside the EU survey, and the nearest comparable evidence comes from the KOF Innovation Survey, run by the KOF Swiss Economic Institute at ETH Zürich on behalf of the State Secretariat for Education, Research and Innovation. Reporting in February 2025 on the period from 2020 to 2022, KOF found just over 8% of small Swiss firms using AI, while for large companies the figure was already more than one in three. Big data showed the same shape — 20% of small firms against 60% of large ones — and fewer than 5% of firms combined the two by using AI to evaluate and analyse large volumes of data.
The levels are not comparable with Eurostat’s: different years, different questionnaires, different populations, and Swiss adoption will have risen since 2022 as European adoption has. The gradient is comparable, and it is the same gradient.
One further KOF observation belongs in any Swiss discussion of budgets. Examining the surge in digital technology during the pandemic, the institute concluded that it was driven primarily by greater employee expertise rather than by higher investment in hardware and software, with ICT spending holding at around 17% of total investment. Approached from the input side, that finding matches what EU firms report from the barrier side. What moved the line was competence, not capital expenditure.
A management team can act on this without a large-firm budget
Nothing in these statistics suggests that firms on the near side of the divide bought better technology. The decisions that separate them are ordinary, and available at any size.
- Choose one process rather than one tool. The European data shows where AI is actually pointed, and where small firms diverge from large ones: 19.02% of small AI users apply it to production processes against 33.46% of large ones, and 14.51% to ICT security against 47.51%. The operational uses are where the distance is widest.
- Budget hours before licences. Training is the input a thirty-person company can buy in small quantities — two people, one process, a fixed number of hours each week over a quarter — and the training statistics describe the gap more accurately than any spending figure.
- Settle the legal and data question once, in writing. More than half of the firms that held back cited legal uncertainty and nearly half data protection. A one-page internal rule stating what may be entered into which system, and who authorises exceptions, converts a recurring hesitation into a decided matter.
- Do not plan to recruit the capability. With 57.5% of recruiting enterprises reporting hard-to-fill vacancies, the realistic route for a smaller firm is to develop the staff it already has.
- Fix the measurement before starting. One number — hours spent on a task, quotation turnaround, error rate — recorded before and after, will settle within a quarter whether to extend the application or stop it.
The next survey rounds will show whether the distance closes
Eurostat runs the survey on ICT usage and e-commerce in enterprises annually and publishes the artificial intelligence results in the isoc_eb_ai dataset; the next round will show whether the small-firm rate closes on the large-firm rate or simply tracks it upward. On the Swiss side, the next wave of the KOF Innovation Survey will be the first to cover a period in which generative tools were broadly available to firms of every size.
Two readings of the present data are both defensible. The divide is real and it accumulates, because experience compounds and the firms acquiring it now begin each following year further ahead. And the cost of entry has rarely been lower, since the technologies most used across Europe today are language tools sold by subscription. What distinguishes the two readings is not the budget a company has, but whether it treats competence as something to be scheduled.