Six Global Agencies Mapped AI Skills. The Gap Is Not Where You Think

28% of European adult workers already use AI on the job. Only about 15% received any training on how to use it in 2023-24. That gap — between how fast AI arrived at work and how slowly anyone is being taught to handle it — is the quiet headline of a report that six international agencies just signed together.

"Changing Landscape of Skills in the Age of AI," published in August 2026, is a joint product of the ETF, Cedefop, Eurofound, the European Commission, the ILO and UNESCO — the Inter-Agency Group on TVET. Most AI-and-work research asks which occupations are exposed. This 60-page review asks a different question: which skills inside your job are gaining or losing value. That is a third layer of the picture. We have covered occupation-level exposure (ILO's ASEAN study) and task-level adoption (the first US task-level survey). Skills are the unit you can actually train.

Nobody can agree how big the "AI workforce" is — and that matters

Start with the number everyone gets wrong. [Fact] Depending on method and definition, the share of workers who develop or professionally use AI ranges from 0.3% to 5% of total employment across the studies the report compares. [Estimate] That is roughly a 17-fold spread between the low and high ends — the measurement instruments disagree with each other by an order of magnitude, which is worth remembering the next time a single headline number about "AI jobs" crosses your feed.

Even the high end is small. [Fact] The IMF finds AI skills appeared in almost 5% of US job postings in 2025, up from under 1% before 2015 — and half of those ask only for AI-user skills, not developer skills. [Fact] Across 30 European countries, AI vacancies average about 1% of all postings; the UK (168,000), Germany (102,000) and France (88,000) together account for 57% of Europe's total.

The more useful finding is where the shortage sits. A study the report highlights sorts AI skills into three tiers — basic (AI-literate users), intermediate (software and data proficiency), advanced (AI research and engineering). [Fact] Intermediate skills are the backbone: 47% of AI-mentioning vacancies and 63% of AI-mentioning CVs sit in that middle tier, broadly matched. The imbalance is at both ends: basic skills appear in 29% of vacancies but only 22% of CVs; advanced skills in 24% of vacancies but only 15% of CVs. [Estimate] Run the division and advanced-tier supply covers roughly 63% of expressed demand, basic-tier about 76% — Europe is short of experts and short of ordinary literate users, at the same time.

Here is the counter-narrative buried in that arithmetic. The standard advice — "learn to code, become an AI engineer" — targets the smallest tier. [Claim] The report's own synthesis suggests the biggest volume of need is at the bottom of the pyramid: basic AI literacy for nearly the whole workforce, not engineering for the few. If you work as a customer service representative or a financial analyst, the skills question is not whether you can build a model. It is whether you can judge one.

What actually changes inside a job

The report's core mechanism is a shift in which cognitive skills earn their keep. [Fact] In Cedefop's 2024 survey, about one-third of European workers no longer do some tasks since AI arrived, roughly 4 in 10 do new or different tasks, and 67% do some tasks faster. [Fact] In OECD case studies of finance and manufacturing, 66% and 72% of employers respectively say AI automated tasks workers used to do — and about half say it created tasks that did not exist before.

What fills the freed-up time is oversight. The report leans on David Autor's framing: critical thinking is moving from gathering information to verifying it, from solving problems to integrating AI output, from executing tasks to supervising them. There is even a word now for what happens when that verification is skipped — "workslop," AI-generated content that looks professionally finished but shifts the cognitive burden onto whoever receives it.

The within-occupation numbers make this concrete. [Fact] In the ILO's task-level automation scores, a primary school teacher's lesson-planning task scores 0.465 in automation potential, while maintaining classroom discipline scores 0.16. [Estimate] That is a 2.9-fold difference inside a single job — the job survives, the task mix shifts. The same logic applies to elementary school teachers, registered nurses and surgeons: AI takes the drafting and the monitoring; humans keep the judgement, the empathy and the exceptions. [Fact] The WEF estimates tasks tied to empathy, creativity and leadership carry only a 13% potential for AI transformation.

One wrinkle the report does not smooth over: the soft-skills story is not uniform. [Fact] In an OECD survey of managers using algorithmic tools, US employers report rising demand for social skills — while European managers report a sizeable decline. "Socio-emotional skills always win" is a comforting slogan; the data says it depends on how your employer redesigns the work.

And a European occupational analysis in the report sorts AI-mentioning job titles into three functions — build, redesign, improve. The build column is the expected cast: software developers and data scientists. But the redesign and improve columns hold translators, journalists, graphic and product designers, marketing professionals — occupations where AI shows up not as a career but as a renovation.

The AI most workers meet is the one managing them

Here is the report's least-discussed number. [Fact] In the US, roughly 90% of surveyed employers have adopted at least one algorithmic tool to instruct, monitor or evaluate workers. The average is 79% across France, Germany, Italy and Spain, and 40% in Japan. [Fact] For 24% of EU workers, algorithms already determine workflow or task priorities; 16-17% have algorithms monitoring or evaluating their performance.

Compare that with usage: about 28% of European workers use AI tools themselves. [Estimate] In the US at least, the share of workplaces where an algorithm manages people appears to exceed — by a wide margin — the share of workers who actively use AI. For many people the first workplace AI is not a chatbot they open; it is a scheduler they never see. That reframes what "AI skills" means: understanding how you are measured is becoming as practical a skill as prompting.

The training system is losing the race

The supply side is the report's bleakest section. [Fact] Between 40% and 60% of surveyed European adult workers have a limited understanding of the basic pillars of AI literacy. [Fact] In a four-country OECD analysis (Australia, Germany, Singapore, the US), only 0.3% to 5.5% of training courses on offer deliver any AI content at all. [Fact] Meanwhile only 25% of US workers were deploying AI on the job in 2024, against roughly 60% in China and India — a gap the report links partly to trust (47% of Americans say they trust AI, versus 77% in China).

Some systems are moving. [Fact] China now requires at least eight hours of AI education annually from elementary school. Estonia partners with AI labs to put tools in front of students and teachers. Finland's free Elements of AI course dates back to 2018. But the report's blunt reading is that current training supply "may not be sufficient to meet demand" — and the people least likely to get AI training are low-skilled workers in the jobs most exposed to automation. The training follows the privilege, not the risk.

What this means for your next twelve months

The practical translation is unglamorous. First, the highest-return skill named across nearly every study here is verification: checking, contextualising and challenging AI output in your own domain. That requires domain expertise — AI literacy is a complement to what you already know, not a substitute. Second, if formal training is not coming (and statistically it is not), the report notes most workers learn AI tools on the job; deliberate practice with the tools your sector actually uses beats a generic certificate. Third, ask how algorithms are used to manage work at your employer — that is an AI-skills question too, and in Europe it is one your works council can raise.

The honest caveats: this is a synthesis of existing studies, not new primary data — its numbers come from surveys with different samples, years and definitions, which is exactly why the AI-workforce estimate spans 0.3% to 5%. Vacancy-based skill measures also systematically under-count basic skills, because employers stop listing what they assume everyone has — the report says so itself. And the ILO task scores it cites belong to the same methodological family as the exposure metrics used on this site, so the limits of one are the limits of the other.

Six agencies rarely agree on anything. On this they do: the scarce resource of the AI transition is not compute. It is judgement, and nobody is funding its training at scale.

Sources

• ETF, Cedefop, Eurofound, European Commission, ILO, UNESCO (Inter-Agency Group on TVET), "Changing landscape of skills in the age of AI," August 2026: https://www.ilo.org/publications/changing-landscape-skills-age-ai
• Full report PDF (60 pages): https://www.ilo.org/sites/default/files/2026-08/Final%20version_IAG%20paper_AI%27s%20impact%20on%20skills%20demand_0.pdf

This article was produced with AI assistance and reviewed by a human editor. All figures were verified against the full text of the source report; derived calculations are marked as estimates.