Europe’s AI training gap is a format problem as well as a volume problem

Adoption has moved faster than flexible supply

The Joint Research Centre published the public record for AI and Skills on 15 October. Its abstract says occupational exposure is broad, actual use lags exposure, about 30% of EU workers reported using AI at work in 2024–25, and about 20% of enterprises had adopted AI by 2025. It also reports no clear labour-market disruption in the employment and vacancy evidence it reviewed.

The numbers are not one adoption rate. Worker use, enterprise adoption and occupational exposure measure different populations and questions. The full report is marked as visible only at Commission level, so this draft does not claim to have inspected its tables or reproduce restricted content. It relies on the public record and triangulates the broad direction with the Commission's employment chapter.

The public JRC abstract identifies a supply issue that deserves operational attention: advanced AI education is concentrated in master's programmes, while short, flexible formats for people already in work are underdeveloped. That is a different problem from simply counting AI courses. People in employment face time, sequencing and employer-support constraints that a degree-shaped offer may not solve.

Measure a format's reach and transfer separately

The OECD's June synthesis says fewer than 1% of workers need advanced AI-specific skills, while most need digital, data, managerial and human capabilities. It also reports that more than half of workers using AI received employer-funded training and that trained users were more likely to report positive outcomes. Those are survey associations across defined samples, not proof that any short course causes better performance.

For a learning portfolio, separate three decisions. First, who needs specialist model-development depth versus applied use, verification or management? Second, which delivery format can that population actually complete? Third, does learning transfer into a changed work practice with acceptable quality and risk?

We propose a format matrix with audience, prerequisite, learning hours, synchronous burden, completion, assessment type, workplace task, manager support and a delayed transfer check. Report access and completion before claiming skill gain. For applied pathways, use a bounded work sample or decision exercise; for advanced pathways, require deeper technical evidence. Compare completion and transfer by employment status, firm size and schedule where lawful and adequately powered. Do not infer that a microcredential is effective because it is short.

Run the first comparison as a bounded service test rather than a portfolio-wide verdict. Offer the same task and support goal through two feasible formats, document who is excluded by schedule or prerequisite, and predefine what counts as usable evidence. Follow up after a fixed interval with one observable work task and a manager or peer check. A difference in completion is not automatically a difference in capability; a difference in self-report is not automatically transfer. Record both null and adverse results so a convenient format is not retained only because its failures disappeared from the dashboard.

The earlier Cedefop self-report article cautioned that perceived preparedness is not observed capability. The new JRC synthesis adds a supply-side question: even a sound curriculum can miss employed learners if the format does not fit their constraints. The immediate action is to map the missing format and its target task, then test access, evidence of capability and workplace transfer as separate outcomes.