AI in public employment services: Unlocking potential, avoiding pitfalls

Public employment services are rapidly turning to artificial intelligence (AI) to develop a range of sophisticated digital tools. According to a recent survey, half of the public employment agencies in OECD countries have implemented AI solutions such as chatbots that assist with service inquiries, profiling models to assess jobseeker needs, tools that guide job-search strategies, and job matching systems. AI can help these agencies optimize resources and improve service delivery. However, implementation remains challenging.

We recently completed a project in cooperation with the European Commission, funded by the European Union via the Technical Support Instrument, developing a jobseeker profiling model based on machine learning for the Greek public employment service (DYPA).  While this technical assistance focused on one application and one AI methodology, it illustrated the potential – and pitfalls – of applying AI to the operations of public employment agencies.

Using AI to profile jobseekers

Profiling models are used by virtually all public employment agencies to estimate the intensity of services that each jobseeker is likely to need. This is done by predicting a jobseeker’s “distance from the labor market”, based on personal characteristics, employment history, and the labor market situation. These models can help with a chronic problem that almost all agencies face - how to allocate scarce resources across a large pool of jobseekers. DYPA is no exception - its limited resources, along with a very high caseload, make it difficult to provide the intensive support that unemployed workers often need. Over half of jobseekers never meet a counselor and when they do, meetings may last for less than 15 minutes.

In addition to strengthening service delivery by public employment agencies, profiling models can also serve as a valuable resource for policy-making entities such as the Greece’s Unit of Experts in Employment, Social Insurance, Welfare and Social Affairs. These models can not only identify which groups are likely to remain in long-term unemployment but, when combined with administrative data, they can provide evidence on which interventions lead to successful transitions for different types of jobseekers.

Several European countries are now using AI to profile jobseekers, moving beyond traditional econometric models. The machine learning model we developed in Greece was effective in predicting unemployment duration, capturing complex patterns that traditional models miss. However, its ultimate success hinges on overcoming key challenges when introducing AI tools, such as ensuring transparency, addressing staff and client resistance, and continuous monitoring. For DYPA, three hurdles stood out - navigating AI’s legal and ethical concerns, assembling quality data, and tailoring the model to fit the agency’s operational needs.

Ethical and legal challenges confronting AI in public employment services

AI is still uncharted territory for many governments, with legal and ethical challenges that are tough to navigate – including around bias and discrimination, data privacy, accountability, and the protection of individual rights.

Even in countries with sophisticated public employment services, missteps have occurred. For example, Austria’s service faced backlash when a chatbot was accused of discriminating against women in providing information on training and career orientation to jobseekers.

In addressing these challenges, France stands out for its strong legal and ethical framework, with clear guidelines that govern the use of AI, dedicated oversight teams, and capacity-building initiatives to ensure expertise.

AI tools are only as effective as the quality of the data they rely on

Governments generate vast amounts of data every day through their administrative processes that can support the development of effective AI tools. However, making data usable for machine learning is no simple task, as ensuring quality and linking databases properly can be highly complex. This was certainly a challenge in our engagement. Extracting, interpreting, and connecting multiple rich yet poorly documented databases from DYPA and other government agencies required intensive collaboration between operational and technical teams.

The need for coordination

In the end, the job profiling project underlined the need for close coordination between the technical team that is developing the tools and the agency personnel who will be using them. Yet this is not always an easy partnership.

At different stages, DYPA’s operational staff asked for a profiling model with features that were not feasible with a machine learning approach and available data. At other times, the AI experts proposed models that were technically sound but did not fully meet DYPA’s business needs. Numerous discussions across these two divides were needed – and in fact will continue to be needed during the implementation stage and beyond - to ensure the models are effectively tailored to meet current operational needs.

Lessons for public employment agencies

While our machine-learning profiling project focused on one specific application, it offered more general lessons on how public employment agencies can realize the potential of AI while avoiding the pitfalls. AI can be a powerful tool to enhance employment services, but its impact will ultimately be shaped by how well agencies navigate the complexities of data, ethics, and human expertise.

The challenge now is not just to build smarter models, but to ensure they work in practice through human oversight - helping jobseekers, supporting counselors, and strengthening the labor market as a whole.

Lessons learned in an EU member country like Greece can help less developed countries where AI implementation faces steeper hurdles due to weak digital infrastructure, limited data availability, and low institutional capacity.