Europe found an AI skills gap. Train the exposed worker, not the average

THE SIGNAL IN ONE SENTENCE

Europe does not have one AI skills gap. It has a collection of mismatches wearing the same name. Some workers use AI every day. Some perform tasks that current systems could affect but have never touched the tools. Some need help using a system inside their existing job. Others may need a route into a different occupation. Small employers may want to adopt useful technology but lack the money, time, data and technical staff to make it work. A generic course count makes all of those people look conveniently similar. The European Commission's Directorate-General for Employment, Social Affairs and Inclusion published a summary on October 23 of a new Joint Research Centre brief called AI and Skills. The JRC record is dated October 15, and the EU Publications Office says the publication was released on its website on October 14. The brief separates three things that are often mashed together: exposure, use and employment effects. Exposure means a job contains tasks that overlap with what current AI systems can do. It shows where change is technically possible. It does not prove that an employer will adopt a system, that the system will work, that a task will disappear or that a worker will lose a job. Use is different. The JRC says 30 percent of EU workers reported using AI at work in 2024 and 2025. Adoption is uneven, reaching close to 40 percent among white-collar workers but only 6 percent among workers in elementary occupations. Use is concentrated in large companies, digitally intensive sectors and northern member states. The enterprise figures need an asterisk. The JRC publication record says 20 percent of enterprises had adopted AI technologies by 2025. The Commission's October 23 summary says about 30 percent of EU enterprises with ten or more employees have adopted AI. Those statements may use different samples, definitions or reporting periods. The reviewed public pages do not reconcile them, so neither number should be passed around as the single European adoption rate without its source and denominator. Employment effects are the third category. The Commission says that through 2025 it found no clear link in Europe between AI exposure and unemployment, earnings or working hours. The JRC record says employment and vacancy data show no clear sign of labor-market disruption so far. The brief explicitly does not forecast an imminent wave of either job destruction or job creation. That is not a finding that nothing will happen. It is a finding that the available evidence does not support a confident aggregate prediction yet. The plain signal is this: Europe should train the exposed worker, not the statistical average. Start with the task, not the job title. A physician, veterinarian or professional-services manager may have high AI exposure because the job contains research, writing, classification or documentation. Those occupations also depend heavily on judgment, responsibility and human interaction. The Commission says they are more likely to be enhanced by AI and may need upskilling inside the same occupation. A general office clerk, keyboard operator or worker in some finance roles may also have high exposure, but a larger share of the job may consist of tasks a system can directly perform. The Commission says those roles face greater replacement risk and may need reskilling into different occupations. That distinction is more useful than announcing that office work is exposed. Training should begin with a task inventory. What does the person actually do during a week? Which tasks require factual accuracy, privacy, physical skill, interpersonal trust, legal responsibility or local knowledge? Which tasks are repetitive? Which tasks create the evidence that another worker later relies on? Which tasks have already changed because an employer introduced a system quietly? Then test the likely form of change. The first path is augmentation. The worker keeps the occupation but gains a tool. Training should cover the exact system, the relevant data, common failure modes, review standards, escalation and the point at which the worker must refuse the output. The outcome should be better work, not merely faster clicking. The second path is task redesign. The job remains, but the bundle changes. A clerical worker may lose routine transcription and gain exception handling, customer explanation, quality checks or coordination. Training must cover the new tasks, and the employer should update authority, workload, pay and performance measures to match. The third path is occupational transition. Enough of the old role changes that the worker needs a credible bridge into different work. That requires more than a short tool demo. It needs a mapped destination, recognition of existing skills, paid learning time, income support, placement help and evidence that the next role is durable. The support test should ask which path a worker is on before assigning a course. Europe's current training supply does not fit that problem neatly. The JRC says AI-related higher education remains concentrated in master's degrees. Those programs can develop specialists. They are not a practical answer for every 43-year-old office worker, hotel employee, logistics coordinator or small-business manager whose tasks may change this year. The short, flexible formats needed by people who already have jobs remain underdeveloped, according to the brief. Short does not have to mean shallow. A good module can be narrow and serious: checking AI-generated invoices against source records, recognizing fabrication in an insurance workflow, protecting client data in a legal office, escalating a machine recommendation in a warehouse, or testing whether a translated customer message preserved its meaning. The module should use the worker's actual tools and policies, include supervised practice and end with an assessment tied to real work. Flexibility cannot mean pushing the course into unpaid evenings. Track who receives paid time, who gets schedule relief, who has access to equipment, who can participate in a local language and who receives support for disability or caregiving needs. A catalog of micro-credentials is not accessible simply because it fits on a phone. Mid-career workers need special attention because they may be highly exposed while sitting far from the education system. They also carry deep knowledge that generic AI programs tend to ignore. An experienced administrator knows where records fail, which exception is dangerous and which policy sentence does not survive contact with a real customer. A mechanic knows what an unusual sound means before a sensor produces a code. A cleaner knows the difference between a tidy room and an infection-control problem. Training should add technical capability to that knowledge, not treat the person as an empty account waiting for a login. Women also need to be visible in the design. The Commission says women are heavily represented just below the highest exposure band, especially in clerical and secretarial work. An average participation rate can hide whether those workers were offered only basic tool training while higher-paid employees received technical, supervisory or transition opportunities. Publish participation, completion, placement, pay and promotion results by occupation, gender, age, education, region, disability, contract type and company size. Protect personal privacy, but make the distribution visible. Small and medium-sized companies are another weak point. The Commission says use is strongest in large, highly digital firms. That makes sense. A large company can hire specialists, negotiate vendor support, clean data, run a pilot and absorb a failure. A local garage, farm, hotel, cleaning company or logistics operator may not have any of those cushions. Finland's National AI Upskilling Initiative offers one example. The Commission says it has delivered more than 2,000 practical AI training sessions to small and medium-sized businesses in fields including automotive services, agriculture, logistics, tourism and cleaning. That is a delivery count, not an outcome result. The next publication should show which businesses participated, which workers received training, what tasks changed, whether the tools remained in use, what time and money the employer invested, whether output or quality improved, and what happened to hours, pay and staffing. Two thousand sessions can describe an impressive effort while saying very little about the people inside it. The same discipline should apply to the Pact for Skills, the Artificial Intelligence Skills Alliance, sector pilots and micro-credentials. Record the problem each program is meant to solve. Name the target occupations and regions. Publish the eligibility rules, provider, curriculum, duration, paid time, employer obligations and cost per participant. Follow the worker after the certificate. Did the person use the skill? Move to a better task? Keep the job? Change occupations? Gain pay? Take on more unpaid review work? Review burden deserves a line of its own. An AI tool can produce a first draft quickly and still increase the worker's cognitive load. Someone must check facts, identify subtle errors, protect confidential data, explain decisions and accept responsibility. If the organization measures only time to first output, the human labor at the end disappears. Training programs should measure how much review a safe workflow requires and whether workers receive time, authority and pay for it. A worker who becomes the unofficial error catcher has gained responsibility, even if the job description stays still. Teachers sit on both sides of the gap. The Commission summary says only about one-third of national school curricula include AI-related competence and around 30 percent of teachers have participated in AI-focused training. A related Commission education summary puts the surveyed teacher figure at 29.7 percent. Those numbers describe participation, not readiness. Teacher training should test whether educators can protect student data, select appropriate tools, detect unreliable output, preserve the learning task, design non-AI alternatives and explain their choices to students and families. Schools also need workload, technical support, procurement rules and leadership. A trained teacher inside an unprepared institution is a person carrying the whole system in a tote bag. Vocational education needs the same systems view. If an occupation changes, the curriculum, equipment, instructors, placements, assessment and employer relationships may all need to change together. Teaching a new interface on old equipment does not modernize a profession. Neither does adding the letters AI to a course that still prepares students for yesterday's task bundle. Europe's policy problem is therefore not to maximize the number of people who can say they completed AI training. It is to find the workers whose tasks are changing, identify what kind of change they face, deliver support they can actually reach and measure whether the result improved their work. That requires better labor-market monitoring. Exposure measures can identify where to look. Worker and employer surveys can reveal adoption. Administrative data can show jobs, hours and earnings. Workplace studies can show changing tasks, review burden, autonomy and safety. Training records can show who received support. None of those sources is sufficient alone. Link them carefully, preserve privacy and publish the limitations. Europe has not found a single skills gap waiting to be filled with a single course. It has found a targeting problem. The honest test is whether the person most exposed to change receives the support that fits their actual path before the path closes.