Europe’s AI adoption count needs a core-process denominator

The Joint Research Centre’s 14 October release reports that the share of firms using AI across ten EU strategic sectors rose from 7% in 2021 to 18% in 2025. It gives 20% for the wider economy in 2025. The release distinguishes general-purpose tool uptake from integration into production and other core processes.

This is a new publication about an earlier observation window, not an October 2026 census of deployments. The underlying report record describes a methodology mapping strategic sectors to Eurostat enterprise ICT data. Its public record and release were inspected here; the complete report was not independently reanalysed.

Keep the denominator attached

Eurostat’s survey definition normally covers enterprises with at least ten employed people in selected economic activities; smaller enterprises are optional. That prevents reading the figure as the share of all employers, workers or tasks using AI. JRC and Eurostat are also one underlying data lineage here, not two independent replications.

For a workforce plan, retain a denominator chain. Start with covered enterprises, then identify the relevant sector and firm size. Inside an organisation, separately count staff with access, people using a tool for a specified task, and processes where AI is integrated into normal execution. Do not combine these into a single adoption score. Each answers a different budget question.

A team may need basic access and support where use is absent. Where tools are already used, a further introductory course may be less useful than task-specific practice. Where AI influences production, the training question can shift towards exception handling, validation and accountability. These are planning hypotheses to test locally, not effects estimated by the JRC analysis.

Correlation does not select the intervention

The JRC reports a positive association between organisational change and production use. It does not establish that reorganisation caused better outcomes. More capable firms could both reorganise and deploy, while sector technology fit could affect each. The headline percentage cannot tell a buyer whether to purchase a platform, hire specialists or redesign a workflow.

Require a process-level baseline before making that choice. Record the current output, error types, waiting time and human effort for a named task. Identify which limitation training could realistically address and which instead requires data access, a better interface or a clear decision right. A skills gap should not become the default explanation for every blocked deployment.

A useful small pilot pairs a learning measure with an operating measure. For example, evaluate whether staff can detect a seeded exception without help, and whether the same process maintains quality with the proposed tool. Keep unsuccessful attempts and abandoned use cases in the record. Counting only active users could hide where the operating model failed.

The newsroom’s earlier analysis of workflow depth offers a related distinction, but the new EU figures should retain their own geography, activity classification and 2021–2025 window. Neither dataset should be turned into a cross-country productivity ranking.

The immediate decision is modest: replace an undifferentiated adoption target with separate measures for access, task use and core-process integration. Then fund learning against the observed constraint. That approach respects the official evidence while avoiding a claim the release cannot support—that more AI use, by itself, proves more organisational capability.