61% of European workers expect new AI skills. Only 15% trained last year
European workers can see the skills gap coming. In Cedefop's 2024 AI skills survey, 61% said they were fairly or very likely to need new knowledge and skills within five years to cope with AI at work. Only 15% had taken AI training during the previous year. Forty-four percent thought their employer was unlikely to provide it.
More courses will not solve the problem if they teach the wrong thing. A new European Training Foundation report, published in August 2026 and also hosted by the ILO, says AI literacy is not programming, neural-network mechanics, or prompt engineering alone. It is the ability to understand, use, monitor, and critically reflect on AI systems.
Most employees will use or be affected by AI without becoming AI developers. The report cites Cedefop data showing that close to one in three European workers regularly use AI tools or technologies, while only about 4% to 5% of those users develop AI. Training should prepare people to make decisions around AI. It should not pretend that every job is turning into machine learning engineering.
The specialist pipeline is too narrow a model
The ETF review puts the technical AI workforce somewhere between 0.3% and 5% of employment, depending on the country, definition, data source, and period. In a 14-country OECD analysis, vacancies demanding AI skills rose from 0.3% of postings in 2019 to 0.4% in 2022 and never exceeded 1% in any country or year studied. Another US analysis found AI skills in almost 5% of postings in 2025, up from less than 1% before 2015.
These measures are not directly interchangeable. Some count developers, some include professional users, and most infer skills from online job advertisements. They do support one narrower conclusion: demand is growing quickly, but jobs building AI remain a small slice of the labour market.
Meanwhile, Cedefop's first AI skills survey findings say more than a quarter of European adults were already experimenting with AI at work and six in ten employees were susceptible to some AI-related task transformation. The exposure is broad even when specialist hiring is not.
A generic prompt workshop is a poor default. It teaches the visible interface but leaves out the job knowledge needed to judge an answer, the data knowledge needed to question an output, and the authority to stop a bad automated recommendation.
Three literacies, not one bag of tricks
The ETF report separates AI literacy into three connected domains.
Technical literacy is a basic understanding of how machine learning, algorithms, and data-driven systems work. A worker does not need to train a model, but should understand that an output is probabilistic, shaped by its data and instructions, and limited by the system's design.
Analytical and data literacy is the ability to interpret and evaluate outputs in context. The report uses a procurement officer reviewing an AI-generated supplier-risk flag: the skill is not getting a score from the system, but questioning anomalies and investigating before acting. In another job, the same domain might mean checking citations, comparing a generated summary with the source record, or recognizing that an automation has silently dropped cases.
Ethical and societal literacy covers privacy, bias, accountability, fairness, transparency, and sustainability. It also includes understanding how AI can influence behaviour, public debate, and access to services. An ethics slide at the end of a product tutorial is not enough. These questions determine which tools should be used, what data may enter them, who is accountable, and when a person must make the final call.
Training built around these three domains remains portable across products. Interfaces change quickly. Checking evidence, protecting sensitive data, documenting human review, and escalating uncertainty remain useful when the product changes.
A practical workplace curriculum
An employer can turn those domains into a short programme without teaching everyone the same material.
Start with a shared foundation. Every employee who encounters AI should know what the approved systems are, what information must not be entered, where generated content requires disclosure, how to report an incident, and which decisions cannot be delegated. The European Commission's AI literacy practice repository now contains more than 40 company and public-sector examples. It also warns that copying one of them does not automatically prove compliance with Article 4 of the AI Act.
Then train around a real workflow. Give a recruiter an AI-assisted screening case, a developer a generated patch, a customer-support worker a proposed reply, or a finance team a flagged transaction. Ask them to identify the source data, test an error case, explain the output's limits, and state who owns the decision. The exercise should end with evidence, not a clever prompt.
Add progression once the foundation is in place. UNESCO's AI competency framework for teachers uses three levels: acquire, deepen, and create. The names can travel outside education. A new user acquires safe operating habits. A role specialist deepens judgment through cases from the job. A technical or process owner creates and evaluates a workflow, with documented tests and escalation paths.
This is where tools such as ish.chat can be useful as a practice environment. Teams can compare how different models handle the same bounded task, inspect failures, and agree on a review rule. That practice builds comparison and decision skills without tying the course to one model.
Training needs time, authority, and worker voice
The report assigns responsibility beyond the individual worker. Employers and public authorities share responsibility for providing training, time, and support. The authors also call for social dialogue so workers and employers can negotiate training arrangements and connect deployment to job quality, rights, and inclusion.
That matters when AI changes the shape of a job. Cedefop found that 30% of AI users experienced some task destruction, while 41% took on new tasks. For 68%, AI mainly helped them do tasks faster, yet 55% still reported limited productivity gains from adoption. Training cannot repair a badly designed workflow or give a worker authority that management has removed.
Credentials create another trap. The ETF report supports short programmes, modular pathways, and micro-credentials for adults, but warns that fragmented certificates can reduce transparency and quality. It recommends connecting them to occupational standards, national qualification frameworks, and quality assurance. A completion badge should mean more than sitting through a vendor demo.
Our earlier look at algorithmic management showed why worker oversight cannot be reduced to software settings. Our analysis of AI provenance and data workers made the same point from the supply chain: people affected by an AI system belong in its account of responsibility.
Making prompt engineering mandatory will not close the training gap. Workers need enough technical understanding to know what the system is doing, enough analytical skill to challenge its output, and enough institutional backing to act on that judgment. That is AI literacy suited to work, not a course designed around a product menu.
