OpenAI published an EU version of its AI Jobs Transition Framework on June 29, extending a labor-market map it first developed for the United States in April. The European version uses the official ESCO occupation taxonomy and Eurostat employment data to classify jobs by the kind of near-term pressure or opportunity AI could create.
The headline is not that one large share of work is simply “exposed” to AI. OpenAI separates occupations into four buckets: jobs that may grow with AI, jobs with higher near-term automation potential, jobs likely to reorganize, and jobs with less immediate change.
That distinction matters. A job can be affected by AI without disappearing. A workflow can reorganize around new tools while people remain central to delivery. Demand for some occupations can grow if AI lowers costs, expands access, or makes more projects viable.
The framework uses four transition buckets
OpenAI says about 12% of EU employment is in occupations that may grow with AI. These are roles where lower costs or new capabilities could increase demand for the work. About 14% is in occupations with relatively higher near-term automation potential. Another 27% is in occupations likely to reorganize, where AI may change workflows and skill needs even when people remain important. The remaining 47% is in occupations with less immediate change.
Those figures should not be read as a layoff forecast. OpenAI is careful to say the categories are not employment predictions. They are a planning map.
That is the useful part. Labor-market planning is often too blunt for AI. A single exposure score can make a job look risky without explaining whether the likely change is automation, augmentation, demand growth, or a new division of labor. The four-bucket framework gives policymakers and employers a cleaner way to ask what kind of adjustment is likely.
Country differences are part of the story
The framework also shows that EU countries do not have the same occupational mix. OpenAI says Luxembourg, Sweden, and the Netherlands have larger shares of employment in occupations that may grow with AI. Germany, Greece, and Italy have larger shares in occupations classified as higher automation potential.
That is not a claim that one country will win and another will lose. It is a reminder that AI labor impact depends on the structure of actual work. A country with more employment in occupations that require licensing, in-person delivery, public-service institutions, or specialized local knowledge will experience AI differently from a country with more work in easily digitized office tasks.
The European angle is important because OpenAI is using ESCO and Eurostat rather than simply porting a U.S. occupational map into a different labor market. AI model capabilities can spread quickly. Jobs do not. They sit inside institutions, training systems, languages, regulations, and service-delivery habits.
Monitoring may matter more than the first map
The report’s practical recommendation is not only “reskill workers.” OpenAI argues that Europe has occupation, training, vacancy, wage, and official statistical systems that could be connected to measures of AI capability and workplace adoption. The idea is to detect transition pressure before it appears in aggregate labor data.
That is the right level of ambition. By the time headline employment statistics show a large change, firms and workers have already adapted or failed to adapt. Better monitoring could help governments see where an occupation is being reorganized, where hiring is changing, where training needs are shifting, and where AI is expanding demand rather than replacing work.
For employers, the framework also gives a more responsible language for planning. A role in the “reorganize” bucket does not automatically mean headcount reduction. It may mean changing tools, changing skill requirements, changing supervision, or changing how work is handed off between people and software.
The caveat is uncertainty
OpenAI’s framework is still an estimate built from occupation categories, data sources, and assumptions about AI capability. It cannot know exactly how a hospital, school, court system, factory, bank, or public agency will adopt AI. It also cannot capture the political choices that shape labor markets: training budgets, procurement rules, worker protections, professional standards, and public trust.
That caveat does not make the framework useless. It makes it a starting point.
The strongest read is that AI labor analysis is moving away from simple exposure tables. The better question is what type of change each occupation faces, how soon that pressure might appear, and which institutions can see it early enough to respond.
