Summary

• Around 30% of EU workers reported using AI at work in 2024–25, according to Joint Research Centre research.
• Usage approaches 40% among white-collar workers but falls to 6% in elementary occupations.
• Researchers found no clear aggregate relationship between occupational AI exposure and employment losses through 2025, while identifying significant training gaps.

Workers in European professional and office-based occupations are substantially more likely to use artificial intelligence than those in elementary jobs, according to European Commission research that identifies persistent gaps in training, employer resources and practical opportunities to adopt the technology.

Approximately 30% of EU workers reported using AI at work during 2024–25, but usage approached 40% among white-collar employees compared with 6% in elementary occupations. Larger organisations, digitally intensive industries and northern European economies also account for a disproportionate share of adoption.

The findings come from the European Commission’s Joint Research Centre (JRC), which published its AI and Skills report on 15 October. The Commission’s employment directorate highlighted the research on 23 October, drawing attention to the difference between jobs potentially affected by AI and workplaces where employees actually use it.

That difference complicates workforce planning. A technology may be capable of assisting with a particular task, but employees cannot benefit without access to suitable systems, permission to use them and the knowledge needed to judge their outputs. The cost of introducing software and adapting working practices can create additional difficulties for smaller employers.

Exposure is not the same as displacement

Occupations differ in the extent to which their tasks overlap with capabilities demonstrated by AI systems. Professional, managerial, technical and clerical work often involves information processing, written communication and analysis, making those roles relatively exposed to recent developments in generative AI.

However, exposure measures the potential relevance of the technology to workplace activities rather than the probability that an entire occupation will disappear. A professional may use AI to prepare documents or investigate information while remaining responsible for decisions requiring judgement, specialist knowledge and interaction with other people.

The JRC identifies doctors, veterinarians and professional services managers as examples of highly exposed occupations in which human responsibility remains significant. In such roles, useful AI applications may change the distribution of tasks while leaving accountability with trained staff.

Other occupations face a different balance. General office clerks, keyboard operators and certain finance roles undertake activities that may be more susceptible to direct automation, particularly where the work follows established rules and requires limited discretionary judgement.

Even here, the consequences depend on how employers reorganise operations. Automating part of a process might reduce the time required for a task, change the skills expected of employees or alter recruitment without immediately eliminating the surrounding job.

Importantly, the JRC found no clear aggregate relationship between occupational exposure to AI and unemployment, wages or working hours in Europe through 2025. That observation does not prove that individual organisations have experienced no workforce changes, nor does it predict what will happen as adoption becomes more widespread.

Labour market statistics can also be slow to capture emerging effects. Employment decisions reflect demand, investment, regulation and economic conditions alongside the introduction of technology. Isolating the contribution of AI requires more detailed information about the tasks people perform and how those activities change over time.

Training rarely fits established careers

Although education providers have increased their attention to AI, the research finds gaps between formal programmes and the needs of people already working. Much higher education provision is concentrated in master’s degree programmes, while shorter courses that employees could combine with existing responsibilities remain less developed.

That concentration can serve students preparing for specialist technical careers, but a mid-career employee working in finance, logistics or public administration may need practical training related to their existing occupation rather than a lengthy academic qualification.

AI literacy also involves more than learning to enter effective instructions into a chatbot. Employees need to recognise inaccurate outputs, understand appropriate use of confidential information and know when an automated recommendation requires further investigation.

Relevant skills differ between occupations. A financial analyst may need to examine source records and assumptions in an AI-generated model, while an administrator supporting public services must understand privacy, accuracy and the legal consequences of information used in decisions.

Training tied to actual workflows may therefore be more useful than general instruction on individual AI products. Employers also need arrangements for testing systems before staff depend on them and procedures for dealing with errors that appear during routine work.

The JRC found that only about one-third of national school curricula include AI-related competences and around 30% of teachers have participated in AI-focused training. Such gaps affect preparation for future employment as well as the education system’s ability to teach critical use of the technology.

European initiatives including the Pact for Skills, the Artificial Intelligence Skills Alliance and the DigComp 3.0 digital competence framework provide material intended to support education and workforce development. Their practical reach depends on whether training providers and employers can turn common objectives into accessible programmes.

Smaller businesses need practical support

Smaller companies often have fewer resources for evaluating AI software, managing information security and redesigning processes. Even affordable commercial applications can require technical preparation before they can be used reliably with business information.

The Commission highlights Finland’s National AI Upskilling Initiative as an example of practical support directed towards small and medium-sized enterprises. It has delivered more than 2,000 training sessions across activities including agriculture, tourism, logistics, cleaning and automotive services.

These sectors have different operating requirements, making tailored instruction more relevant than a uniform programme centred on software development. A logistics business might investigate demand forecasting or administrative work, while a tourism operator could examine booking information and customer communication.

The reported number of sessions demonstrates programme delivery, not independently measured productivity gains at participating businesses. Evaluating the initiative would also require information about whether employees retained the skills, adopted useful applications and improved processes after training.

The distinction is important when public programmes assess success through enrolments or attendance. Participation can provide a foundation for improved technology use, but the return depends on subsequent behaviour and whether the software is suitable for the work involved.

Better evidence for workforce decisions

The JRC’s findings create a difficult policy choice because the employees most likely to need support are not necessarily those using AI most frequently today. Workers in highly exposed occupations may need to adapt their existing skills, while those facing the prospect of more extensive task substitution could require preparation for different roles.

Employers similarly need evidence about the technology’s actual effects before making decisions about recruitment or restructuring. A demonstration that AI can complete an isolated task does not establish how it will perform within a larger process, including the checks and corrections required from employees.

More detailed occupational data would help researchers distinguish changes in task allocation from outright job displacement. It would also enable training providers to target activities where new technical skills can be applied immediately.

For now, the research demonstrates a substantial difference between the reach of AI capabilities and their use in European workplaces. Closing that difference will depend on suitable training, credible information about results and employers’ ability to integrate technology without overlooking the judgement and responsibilities that remain with their staff.