Eurostat’s AI benchmark: adoption is rising, but it is not autonomous buying

External-source research synthesis, not an OfferSignal client result. Based on the retained October 2026 research and source-specific dates.

A business deciding whether to invest in AI needs more than a headline adoption number. “AI use” can mean text analysis, image generation, workflow assistance or machine learning; it does not necessarily mean that a buyer discovers suppliers, approves a quote or places an order. Eurostat’s 2025 ICT-usage survey is valuable precisely because it makes the denominator and the limits visible.

What Eurostat measured

Eurostat reports that 19.95% of covered EU enterprises used at least one listed AI technology in the 2025 survey year, up 6.47 percentage points from 2024. The population is enterprises with at least 10 employees and self-employed persons in covered activities; micro-enterprises below 10 are optional and are not part of the mandatory headline population. Financial activities and several primary-sector activities are excluded. The observation unit is an enterprise, not a transaction, customer, employee, quote or order. 1 2 3

The listed technologies include text mining, speech recognition, natural-language generation or speech synthesis, image recognition, machine learning, AI-based workflow automation or decision assistance, autonomous physical movement and, newly in 2025, generation of pictures, video and sound/audio. The aggregate therefore represents enterprises using at least one technology, not a measure of autonomous procurement or deployment quality. 1 2

Size, country and sector change the picture

The size gradient is substantial: AI use was 17.00% among small enterprises with 10–49 persons employed, 30.36% among medium enterprises with 50–249, and 55.03% among large enterprises with 250 or more. That is useful for segmentation, but “small or medium enterprise” must not be treated as a synonym for a business-to-business buyer or supplier. 1

Country shares ranged from 5.21% in Romania to 42.03% in Denmark. Finland was at 37.82%, Sweden at 35.04%, Belgium at 34.54% and Luxembourg at 33.61%; Poland was 8.36% and Bulgaria 8.55%. The exact Estonia API observation was 23.40% of covered enterprises with at least 10 persons employed using at least one AI technology in 2025. These comparisons can support market-sizing hypotheses, but sector mix, enterprise-size mix, implementation, non-response and reliability differences limit causal interpretation. 1 2 4

Sector shares were highest in information and communication at 62.52% and professional, scientific and technical activities at 40.43%, compared with 10.79% in construction. Among individual technology categories, written-language analysis reached 11.75%, generation of pictures/video/audio 9.55% and generation of written or spoken language or programming code 8.76%; autonomous physical movement was 1.39%. Among AI-using enterprises, 34.70% reported marketing or sales use, 31.05% organisation of administration or management, and 6.08% logistics. These are prevalence and purpose signals, not evidence of a complete commercial workflow. 1

Adoption friction is not purchase intent

Among enterprises not using the listed technologies, 14.21% had considered using one: 12.65% of small, 22.26% of medium and 36.54% of large enterprises. Within enterprises that had considered AI but did not use it, the leading reported barriers were lack of relevant expertise at 70.89%, lack of clarity about legal consequences at 52.52% and data-protection or privacy concerns at 48.83%. Only 20.68% said AI was not useful. These conditional percentages do not establish buying intent or show that a particular product resolves the barrier. 1

Eurostat says the 2025 survey covered about 157,000 of approximately 1.53 million EU enterprises, with roughly 83% small, 14% medium and 3% large. The survey is strong for official, comparable segmentation, but weak for causal business outcomes. It does not link adoption to productivity, revenue, margin, cycle time, accepted offers, safety or realised savings. The 2025 generative pictures/video/audio category is new, so it has no 2024 comparison. 1 2 3

An illustrative workshop test

For an illustrative custom manufacturer, use the benchmark to choose interview and experiment strata—not to claim demand for an autonomous purchasing agent. Test one governed workflow: retrieving the correct drawing or project record, checking revision and units, confirming material and tolerance compatibility, checking equipment capability and current capacity, and preparing an assumption-tagged draft for review. The authorised engineer or commercial owner remains responsible for feasibility, price, terms, quote release and any order.

Measure record-retrieval precision, wrong-revision rate, constraint completeness, human approval rate, quote cycle time, accepted-offer rate and realised margin. Keep source and version controls, and treat language, value-added tax, Incoterms, duties, sanctions, export controls and certificates as separate operational evidence. Eurostat’s 19.95% EU or 23.40% Estonia prevalence cannot validate any of those decisions.

Sources and next reading

Eurostat AI-by-size dataset · AI-by-activity dataset · dissemination API

Explore OfferSignal · Focused Business & AI Workshop · All research

References