This chapter intends to connect the design of the proposed AI-enabled job matching and career recommendation platform discussed under the previous chapter to the practical steps needed for its implementation. It highlights key complementary considerations that underpin the successful adoption of these tools, including organisational and data requirements, pilot testing and monitoring mechanisms, as well as the principal constraints, risks, and challenges arising from the deployment of advanced AI functionalities. These issues are particularly important because eventually the effectiveness of AI-driven matching and career recommendations solutions depends not only on the actual design of the tools themselves, but also on the quality of the underlying data, the supporting infrastructure, as well as their successful integration into the wider ecosystem of employment and social services.
AI and Digitalisation for Employment Support in Belgium and Greece
Abstract
Having discussed the proposed concepts and governance options for the AI-enabled job matching and career recommendation platform in the previous chapter, this chapter focusses on implementation considerations. The successful deployment of such tools requires not only appropriate governance arrangements but also careful planning to address technical, organisational, legal, and operational challenges.
The chapter is organised as follows. It first outlines an indicative implementation roadmap for the proposed job matching and career recommendation platform. It then outlines the main constraints, risks, and challenges associated with deploying advanced AI functionalities and discusses the importance of responsible, human-centred design to ensure that the platform meets user needs and supports rather than replaces, human judgement. The chapter also explores options for piloting the platform and emphasises the need for ongoing monitoring and evaluation to ensure its continuous finetuning as well as complementary policy actions related to organisational readiness, institutional co‑ordination, long-term financial sustainability, availability of good jobs and ALMPs that could strengthen even further the impact of digital tools on MIS beneficiaries’ employment outcomes.
The effectiveness of these tools depends heavily on access to high-quality data and the infrastructure needed to support secure data sharing across different institutions in both countries. Drawing on the data sources, as these have been discussed in Chapter 2, this chapter concludes by providing recommendations on the data that could support the proposed tools in Greece and Belgium and briefly discusses key legal considerations governing the use and sharing of these data.
Table 6.1 presents an indicative implementation plan for the job matching and career recommendations platform discussed under the previous chapter. The plan is organised by phase and outlines the key objectives of each phase, an indicative set of main actions, and estimated human and financial resource needs. Resource requirements are expressed in broad categories (low, medium, high) to give a general indication of the effort required at each phase. It is important to note that the resources related to digitalisation refer to the upfront investment required from the managing authorities during the platform’s development and rollout. Over time, the relevant investments are expected to generate efficiency gains and reduce costs. Only the financial costs of external IT service providers who will be likely contracted for the actual development of the tool are taken into consideration in the estimates below. As far as human resource needs are concerned, these refer to the staff time required within the organisations involved.
Note: The implementation plan will heavily depend on whether the platform is built entirely from scratch or developed by adapting existing tools.
Source: Authors’ elaboration.
Figure 6.1 presents a provisional timeline to show how the different phases unfold over the full implementation period and beyond.
Note: The timeline will heavily depend on whether the platform is built entirely from scratch or developed by adapting existing tools.
Source: Authors’ elaboration.
The implementation of an AI-enabled job matching platform involves several steps, and this section highlights the key issues in this process. It discusses the main risks and safeguards associated with AI deployment, approaches to responsible design, piloting, monitoring and evaluation, as well as broader policy measures.
6.3.1. The AI functionalities of a matching platform may entail several risks and challenges
There are several constraints, risks and challenges associated with the design, development and operation of an AI-powered matching platform. These issues can affect – to a lesser or greater extent – the platform’s performance, user experience, trust, and compliance. The identified constraints, risks and challenges fall into the following seven categories:
1. Data privacy and security issues: The platform must comply with fundamental data protection principles such as the General Data Protection Regulation (GDPR) and the EU AI Act, as it processes personal and sensitive personal information such as employment histories, skills profiles, socio-economic characteristics, and beneficiary data (even if details such as gender, age, or nationality are not meant to be used for matching purposes). Ensuring lawful data processing, consent management, secure storage, encrypted transmission, algorithmic decision transparency, cross-institutional data-sharing arrangements and carefully managed access rights is essential. Unauthorised access or data breaches through for example hacking or weak access controls, could significantly undermine the platform’s credibility and create serious legal implications and reputational risks for the managing authorities.
2. Algorithmic bias and fairness: The underlying AI algorithms may contain biases or discrimination which in turn could perpetuate or amplify existing labour market inequalities. This may lead to unfair treatment of certain groups, including the MIS beneficiaries themselves, women, older workers, or people with disabilities, ultimately harming their employment opportunities and prospects. If for example, certain sectors historically employ more men than women, the platform may at the end suggest to women fewer positions in technical fields. Or if the underlying model uses indicators such as previous employment history, users who have been out of work for longer (which might be the case for MIS beneficiaries) may end up receiving fewer or lower-quality job suggestions.
3. Trust, transparency and accountability: These are key as an AI-powered matching platform can influence and affect not only people’s decisions but also their lives. If users do not understand how the system works, how the results are produced or if they feel they cannot rely on it, they may be hesitant to engage with it or may question its outcomes. Likewise, if it is unclear who is responsible for monitoring the system and addressing issues, it can be difficult to identify and correct potential errors.
4. Accessibility and exclusion: Not all users begin from the same starting point. In practical terms, this means that naturally users may have different levels of digital literacy, or access to digital means and internet connectivity. These discrepancies may affect how easily they can use the platform and limit their ability to benefit from the platform’s functionalities. More broadly, accessibility-related issues present a challenge for the adoption of AI tools in general as there is a risk that the digital divide could become even more pronounced if certain users are unable to engage with the technology effectively.
5. Technical and operational issues: The platform’s technical complexity may entail challenges related to model accuracy, system stability, ongoing maintenance, and the ability to scale as usage grows. Real-time processing of large datasets, connecting with multiple systems, and keeping the system updated to reflect labour market developments all require continuously available and adequate technical resources. Operational issues, such as temporary system downtime, outdated information, or limited technical support may also hinder both user experience and the platform’s overall performance. For instance, if the platform experiences a short period of downtime during peak usage hours, jobseekers may not be able to access their recommended matches, and counsellors may have limited information. Similarly, if labour market data is not updated regularly, the platform may continue suggesting roles that are no longer available, limiting its usefulness for jobseekers and employers alike.
6. Misinterpretation and misuse of results: AI-generated matches offer probabilistic suggestions that are indicative. Some users, however, may mistakenly interpret these suggestions as mandatory, or even linked to their eligibility for benefits or preferential treatment by the managing authorities. Over-reliance on automated results by all types of end users – including jobseekers, employers, counsellors and social workers – may lead to inappropriate referrals or overlook important contextual factors. Likewise, policymakers may draw incorrect conclusions from aggregated results, potentially leading to poorly informed decisions.
7. Data quality and interoperability: The matching platform relies on various information systems, which though are not always adequately interconnected. What is more, data formats, APIs, or synchronisation mechanisms are not always standardised and fully aligned. These issues may result in fragmented information, inconsistent or outdated records, or duplicated administrative effort, all of which can have an adverse impact on the quality and reliability of the AI-driven matching process and other related processes. If for instance, an employer posts a new job vacancy into one of the underlying systems which is not synchronised promptly with the matching platform, the vacancy may not appear on the platform in time. As a result, some jobseekers may miss the opportunity, while the employer may receive fewer applications.
6.3.2. Responsible design and human-centred implementation will ensure the AI-based matching tool delivers inclusive, reliable, and sustainable results
Provided that the aforementioned constraints, risks and challenges are identified at the outset, and in order to fully seize the benefits of AI, it is essential to take proactive steps to mitigate their impact (Brioscú et al., 2024[1]). In doing so, the two countries can draw on established international frameworks that guide the responsible use of AI, such as the OECD’s AI Principles, which seek to promote trustworthy, transparent and accountable AI systems (OECD, 2024[2]).1 Among other important issues, such frameworks highlight the importance of responsible design and human-centred implementation, hereby ensuring that AI solutions are developed with the needs of users in mind, remain aligned with broader public-interest objectives, and support rather than replace human judgement.
The most important steps are the following:
1. Establish a strong governance framework surrounding AI use in the matching platform: A clear governance framework can help ensure that AI is used responsibly, ethically, and in compliance with rules such as GDPR and the EU AI Act. Oversight bodies, such as Ethics Committees or Boards (e.g. PES in France and Flanders in Belgium) can provide regular monitoring, evaluate risks, and guide improvements.
2. Ensure human oversight and allow for flexibility in use: AI should complement, rather than replace, professional judgement. Managing authorities in both Belgium and Greece should make it clear that while users can access and use the platform independently, final decisions remain with staff and that they can override AI recommendations when and as needed. This can help reassure both jobseekers and employers but also the staff of the competent agencies that they still have control and can make their own decisions. This is also aligned with legal requirements such as GDPR (Article 22) which protects individuals by giving them the right not to be subject to decisions made solely through automated processing (Council of the European Union and European Parliament, 2016[3]). Especially when it comes to vulnerable jobseekers such as the MIS beneficiaries, face‑to-face interactions remain critical (OECD, 2021[4]) as some of them may lack the digital skills, may not have access to the necessary digital infrastructure (e.g. internet connection, computer, or smartphone) or need additional support by a human after having interacted with the platform. Finally, the use of the platform should remain voluntary for all end users. Allowing jobseekers and employers to choose whether and how they engage with the tool helps maintain trust and ensures that no one feels obliged to rely on AI-generated recommendations. The same applies to counsellors and social workers.
3. Involve and support end-users early and regularly: Engaging both staff as well as jobseekers and employers from the beginning and keep them engaged on a continuous basis would help ensure the platform is useful, user-friendly, and trusted. Feedback can greatly improve buy-in. It can be indeed collected during the design phase as well as prior to rollout through several ways including consultations, testing procedures and satisfaction surveys. It’s also important that staff who will be interacting with the tool receive relevant support including through training, informational materials, guidelines, as well as a dedicated contact point for technical support. Informational and guiding material for jobseekers and employers should also be available on the platform itself.
4. Collaborate closely with contractors and build internal expertise: More often than not, the development of digital solutions such as the AI-powered matching platforms are outsourced to external contractors. Close and regular communication and collaboration with them is essential to ensure that the actual development occurs as initially planned, tools meet user needs, and that bottlenecks or technical issues are resolved quickly. In some cases, developers work from within the agency, allowing them to better understand the general context and collaborate closely with staff to ensure that technical aspects are developed in line with the operational processes (e.g. PES in Norway and Luxembourg). Rather than relying solely on external developers, a sufficient level of internal expertise vis-à-vis the matching platform should exist within the managing agencies to oversee the tool throughout its lifecycle. This involves understanding how the platform works, monitor its performance, guide its initial development as well as any future updates, ensure long-term sustainability and support other staff as well as the end users (perhaps through a train-the‑trainer approach, where a selected group of staff receive training from the developers).
5. Strengthen trust, transparency, and accountability: Users need to understand how AI-generated recommendations are produced and who is responsible for the results. This includes clear explanations of the data used, the logic behind the model, and the safeguards in place and would help to build trust among the tool’s end users. For example, if a job counsellor does not understand which information the matching platform used or how it generated its suggestions, they may provide guidance that does not fully match the jobseeker’s situation or disregard more suitable opportunities. Furthermore, clear accountability mechanisms including regular risk assessments and monitoring of system performance are needed so that users know who is responsible when errors occur – whether the issue stems from the underlying algorithms, incorrect data, or operational misuse. Transparency and explainability on how the matching platform uses AI should apply both within the managing organisations and towards the public. For instance, the PES in the Netherlands, publishes information about how AI is used, including data sources, bias-mitigation safeguards, and the measures in place for human oversight (UWV, n.d.[5]).
6. Promote inclusion, fairness and responsible use: The matching platform should above all be accessible and straightforward to everyone. This includes supporting staff, jobseekers and employers with limited digital literacy, connectivity or access, including by providing alternative channels for those who cannot use it. Such channels may include for instance multi-device compatibility, offline access, low-data modes, and SMS/USSD/IVR channels as well as in-person support points.2 Enriching the platform with features that can be particularly helpful to the more disadvantaged including people with disabilities and elderly citizens and ensuring that its interfaces meet accessibility standards so that no individuals feel excluded is also key (e.g. interpretation to different languages, sign language, screen reader compatibility, large font options, adjustable contrast, zoom functionality, mobile friendly interfaces, captions/subtitles for audio or video content, visual alerts instead of only sound notifications). Users with low literacy levels can be supported through multiple ways too (e.g. use of icons, visuals, voice guidance, and simplified text). User support and/or training can be provided to these groups to ensure they can use these tools effectively (e.g. with the help of guided onboarding, tooltips and tutorials). To ensure that the AI-driven matching tool is fair, the implementation of measures that address bias is really key. This starts with building fairness into the system from the design stage. Safeguards – such as bias audits, representative datasets, and ongoing monitoring – can help ensure that recommendations remain equitable and do not disproportionally disadvantage vulnerable groups. Finally, clear guidance and appropriate training are essential to help users understand how to use the tool correctly, how to interpret its suggestions, and when to rely on their own judgement, thereby preventing misinterpretation or over-reliance on AI-generated results.
6.3.3. Testing is essential for identifying what works well and allows for refinements before the platform is rolled out
Once the technical and operational details of the matching platform are defined, it is important to test it through a pilot phase before any nationwide implementation. A pilot phase would allow the tool to be used in a real environment, understand how users interact with it, identify what works well and not that well and make improvements before scaling it up or rolling it out nationwide.
There are various approaches to piloting a digital matching platform, depending on the goals, priorities, resources and time constraints of the managing agencies:
Pilot with regional focus: One approach is to introduce the tool in a small number of regions or employment offices. Testing the tool in different areas helps reveal how it works in different local labour markets comprised of different user characteristics and contexts. For example, regions with high unemployment versus areas with increased labour demand.
Pilot with target group focus: Provided that the target group in this very specific case is MIS beneficiaries, the idea would be to focus on this group but with different backgrounds and needs. Testing the tool with this group helps assess whether its specific functionalities (such as skills identification or personalised advice) genuinely support users with more complex needs and barriers. On the employer side, the tool can also be tested with a small number of firms willing to experiment with a new recruitment solution. This can provide insights into how employers use the platform, whether the candidate suggestions are relevant, and how they react to analytics on outreach or diversity.
Pilot with sectoral focus: Another approach involves testing the tool with a focus on specific sectors; for instance, construction, services, hospitality, retail, health and social care. This is useful when sectors face acute labour shortages, require highly specific skills, or have fast-changing job profiles. Sector pilots help verify whether the tool correctly handles sector-specific skills, training requirements, and types of vacancies. Each country can select different sectors depending on the context, priorities and other factors.
Pilot with functionalities focus: A phased pilot can also test only certain functionalities of the tool first. For instance, the recommendation engine, vacancy classification, skills assessment, or labour market information functions could be trialled separately. This approach makes it easier to refine each functionality before integrating everything into a full system.
Regardless of the pilot approach chosen, certain actions should run in parallel. Counsellors and staff need to be trained so they understand the tool and can support users effectively. User engagement should be monitored, including which features are most used, what recommendations the system produces, and whether these suggestions feel relevant to jobseekers and employers. Feedback can be gathered through surveys, interviews, and help-desk queries. Early indicators should also be tracked. For example, whether MIS beneficiaries report feeling better supported, whether the tool helps them discover suitable vacancies they might have missed, or whether employers find the proposed candidates appropriate. As an example, if jobseekers frequently ignore suggested vacancies, this could indicate that recommendations are not well-targeted or that the rationale behind suggestions is not clear enough.
Given that there is limited rigorous evaluation of similar tools internationally, countries have an opportunity to pioneer in this field by integrating an evaluation design from the start. If feasible, the pilot could incorporate a randomised controlled trial (RCT), where access to the tool is randomly assigned across for instance local offices. When properly designed and implemented, an RCT can provide strong evidence on whether the tool improves matching outcomes more effectively than existing services (see also Section 6.3.4 on monitoring and evaluation).
Once the pilot phase is completed, the findings could be then used to refine the tool. This may include improving the user interface, adjusting the matching logic, strengthening guidance features, and addressing any shortcomings identified. After these improvements have been made, the system can be gradually scaled. Depending on the pilot design, this may mean expanding to more regions, involving additional users such as jobseekers, employers or counsellors, integrating further data sources, covering more sectors, or adding on more functionalities. Scaling should take place on a step-by-step basis and remain flexible, so the system can grow without losing quality, stability, or reliability.
6.3.4. Continuous monitoring and evaluation of the tool shall ensure that it operates as intended, improves outcomes for users, and is fine‑tuned based on evidence
Just like ALMPs, digital and AI solutions should also be subject to rigorous and regular monitoring and evaluation to ensure model performance and assess the impact on end-users (OECD, 2022[6]; Brioscú et al., 2024[1]). Evaluation of the matching tool refers to the systematic and objective assessment of the tool’s design, implementation, and results. The purpose of such an evaluation is to determine the relevance of the tool, the extent to which it fulfils its intended objectives, its effectiveness, and its broader impacts. Evaluation results can guide policymakers in making modifications or developing future tools.
A key component of rigorous evaluation is counterfactual impact evaluation, which aims to establish a causal link between an intervention and its intended outcomes. For the matching platform, this means assessing whether the support provided through the tool has a measurable impact on labour market outcomes of its users as well as on any other groups affected by the tool. In practical terms, this involves comparing the employment outcomes of MIS beneficiaries who used the tool (the so-called treatment group) with those of a similar group of beneficiaries who did not use it (the control group), ensuring that any observed differences can be attributed solely to the tool’s existence. For employers, a counterfactual impact evaluation can assess the extent to which the tool improves recruitment outcomes – for example, whether vacancies are filled more efficiently or with better matches than would have occurred without such a tool.
In the evaluations of digital matching tools, typical outcome indicators include: job finding rates, employment status, job retention, job quality (contract type, working hours, job stability), earnings and income effects, occupational mobility, employer satisfaction and vacancy fill rates. Evaluation effects may also extend beyond labour market outcomes to include broader social and health indicators, providing a more comprehensive understanding of the tool’s overall impact, provided the necessary data exist and can be linked appropriately (e.g. deduction in dependency on social assistance, social inclusion, financial well-being, general health status).
With a few exceptions (e.g. the SEND@ digital recommender tool used by the Spanish PES), most digital and AI-powered tools currently in operation have not undergone rigorous impact evaluations (OECD, 2023[7]). Given both the general lack of evaluation evidence and the increasing strategic importance of such tools, there is substantial momentum to incorporate robust counterfactual impact evaluation into their development. In the case of pilot projects, this is particularly valuable for informing decisions on scaling up and national rollout. If a tool is designed from scratch, it also becomes feasible to embed a randomised controlled trial (RCT) within its design. An RCT is considered the most robust method for establishing causality, as random assignment to the tool ensures that any differences in outcomes between participants and non-participants can be attributed to the tool rather than to pre‑existing differences among beneficiaries.
Strengthening the evidence base would not only support more refined improvements in design and implementation of the tool (e.g. identifying which segments of jobseekers benefit most or how counsellors and social workers can use the tool more effectively), but also increase buy-in among users and facilitate broader adoption (e.g. other MIS beneficiaries who were not convinced in using the tool in the first place, other potential users such as jobseekers, employees seeking new opportunities, employers etc.). In the case of both Belgium and Greece, there is already a solid foundation for data collection and exchange to support such evaluations. This includes reliable and good quality administrative data, interoperability between registries and the possibility of integrating additional data sources.
Alongside impact evaluations, process evaluations are also essential. These evaluations can examine whether the matching tool is embedded in service delivery as intended, whether staff and users understand and use it appropriately, and whether implementation challenges exist. Process evaluations are able to collect structured user feedback to understand how useful the tool is for different sub-groups – e.g. individuals with various employment barriers, men vs. women, younger vs. older beneficiaries, and those with disabilities or health conditions. Inputs from job and employer counsellors and social workers is equally important, as they can provide assessments on usability and integration into workflow based on their expertise and valuable experience in serving both jobseekers and employers. These evaluations may rely on focus group discussions, interviews, administrative data, and survey data capturing well-being, motivation, and social integration outcomes.
Cost‑benefit analysis can then determine whether the matching tool generates net benefits once both operational costs and outcome improvements are taken into account. Such analysis can help quantify the tool’s overall value and distributional effects, and it can also be disaggregated to reflect the perspectives of different stakeholders. From the user perspective (e.g. jobseekers and employers), the focus may include time savings, improved matching quality, reduced search costs, as well as any reductions in reliance on social assistance. From the counsellor and social-worker perspective, gains may arise through reduced administrative burden and increased time available for personalised support. It also allows policymakers to compare the tool’s returns with alternative interventions and assess whether scaling up represents an efficient use of resources. From a societal perspective, the analysis can incorporate broader benefits such as reduced dependency on social assistance, increased tax revenues from higher employment, and more efficient functioning of the labour market.
Finally, ongoing and careful monitoring of the tool is critical to track implementation over time. Developing a clear and diverse set of monitoring indicators – covering usage patterns, system performance, data quality, and distributional effects – from the outset and defining the frequency, data sources and responsibility for monitoring will help ensure continuous improvement of the tool and its underlying algorithms.3 Ideally, KPIs should be accompanied by target values. These target values should be derived from analysis, ensuring that they are grounded in realistic and achievable expectations. Ideally, the target values should be set not only at a national level but can also be disaggregated by local employment office or region. As the platform’s use is more frequent and expanded, the initial target values can be adjusted based on continuous performance data, fostering continuous improvement and more accurate goal setting.
6.3.5. Complementary policy actions beyond the introduction of digital matching tools
Digital matching and career recommendation platforms, even when perfectly designed and comprehensive enough to be able to respond to the needs of all stakeholders – whether jobseekers, workers, employers, job counsellors or social workers – cannot fully address all factors that influence employment outcomes. Some of these issues lie beyond the scope of digital tools per se, even though they remain highly relevant to the activation of jobseekers, particularly the ones with weaker labour market prospects. This section touches upon some of these issues. These are framed as policy actions that could be pursued in parallel with the introduction of digital tools in order to further improve employment outcomes.
Four important issues are being discussed. The first one focusses on organisational readiness, co‑ordination and long-term financial sustainability. The second one discusses the importance of good and sustainable employment opportunities for MIS beneficiaries and beyond, including through stronger engagement with employers. The third one is about ensuring the continuous provision of ALMPs and related programmes. The final point concerns reducing or closing information gaps by improving outreach and communication with target populations.
Ensure organisational readiness, co‑ordination and long-term financial sustainability
Most of the times, successful implementation requires not only technical deployment but also wider organisational change. This includes new workflows, updated procedures, and a shift toward data-driven decision making. Resistance from staff, insufficient training, uncertainties around new responsibilities, and disparities in institutional capacity can hinder successful adoption of the matching platform and any other digital tool. Given the number of stakeholders involved in both countries, weak co‑ordination and co‑operation can lead to fragmented decision making and delays in implementation.
Furthermore, building an AI-based matching platform requires long-term financial sustainability and clear budgeting for future needs. This involves not only the investments associated with the tool’s initial design and development but also for ongoing maintenance, updates, training, monitoring, and evaluation. Insufficient or irregular funding may hinder continuous improvement or compromise the quality of the platform.
Expand good and sustainable employment opportunities for MIS beneficiaries and individuals with vulnerabilities
The development and implementation of digital solutions aim to improve matching possibilities for MIS beneficiaries and people in vulnerable situations but also enhance the general conditions surrounding matching to adequately respond to the specific needs of this target group. This approach assumes to a large extent that jobs for people in vulnerable situations exist, but several factors prohibit them for being found and that matching is either completely missing or inadequate. Both of these facts are largely true, and these are issues that can be resolved or at least partially treated as part of the development of the functionalities that have been extensively discussed under the two concepts. For instance, functionalities that can adjust job postings to correct for biases and better align with the profiles of MIS beneficiaries, or functionalities that can effectively identify and showcase the skills and prior experience of this group, so MIS beneficiaries are not overlooked by employers.
Nevertheless, anecdotal evidence suggests that the issue may be deeper. Employment opportunities specifically for people in vulnerable situations are more often than not limited and, when they do exist, their quality and sustainability are inadequate. A common scenario relates to the existence of many candidates, particularly in entry-level or low-skilled roles. When many jobseekers compete for a limited number of opportunities, it leads to frustration, disengagement, and ultimately a loss of trust in the platform and the system in general. For users with vulnerabilities, repeated rejection can have particularly strong negative effects including on their mental health. Despite the fact that labour market conditions in each country are of course different, some of the underlying factors and barriers affecting the availability of jobs for this group may be common cross-border.
One important barrier relates to employer perceptions and stereotypes that are often translated into hiring practices. Some employers may be reluctant to hire individuals from vulnerable groups, including people receiving social assistance, long-term unemployed individuals, migrants, or individuals with limited skills, work experience or work capacity. Such perceptions may stem from the fact that they have limited experience working with such candidates and may therefore be uncertain about their skills, reliability or ability to integrate into the workplace. As a result, the vacancies they post may indirectly exclude these individuals, or they may be filtered out later during the interview or hiring process. In these cases, the challenge for PES staff is not only matching jobseekers with vacancies but most importantly encouraging employers to consider a broader pool of candidates.
Addressing this issue requires comprehensive employer engagement strategies and tailored actions. This includes financial incentives for hiring individuals who have been out of the labour market for an extended period as well as initiatives to support the hiring of individuals from vulnerable backgrounds. For instance, the PES in Sweden provides incentives to employers to hire more jobseekers with disabilities and other hiring obstacles such as the lack of Swedish language skills, through a range of measures, including wage subsidies that intend to offset jobseekers’ reduced productivity (OECD, 2025[8]). Developing programmes – such as training or work experience programmes – in co‑operation with employers can generate multiple benefits. Such programmes can be tailored to employers’ skills needs while providing individuals from vulnerable groups with practical work experience and exposure to real workplace environments. At the same time, the PES can establish arrangements with the participating companies to engage candidates from these groups in those programmes and potentially recruit a share of them upon completion. Other initiatives can include job shadowing (like the “Duo-Days” model introduced by the PES in Slovenia), work trials, immersion days, or company discovery days. These can be particularly valuable for vulnerable jobseekers, while at the same time they can help reduce employer uncertainty prior to hiring. Such initiatives allow employers to observe candidates’ capabilities directly, assess how well they may fit within the team, identify potential adjustments or support needs and help build mutual trust. Finally, workplace adaptation support or dedicated job counsellors can assist both the employer and the employee during the first months of employment.
Some employers may be willing to hire individuals facing vulnerabilities but may hesitate because they feel they won’t be able to provide the necessary support in the workplace. In such cases, the PES could offer practical assistance and guidance to employers. For example, the PES in Estonia provides various forms of workplace support, including funding for necessary equipment for employees with disabilities, adjustments to working arrangements such as reduced working hours, and other forms of support tailored to individual needs. The provision of post-hiring support from the side of the PES is also essential, especially for jobseekers with vulnerabilities. Communicating the availability of such support to employers from an early stage can help build confidence and eliminate any potential concerns. This support may include follow-up check-ins, assistance with workplace integration, guidance in addressing any challenges that may arise, and training where needed to help individuals adapt to the work environment.
Addressing employer perceptions should involve clearly articulating the benefits of inclusive recruitment practices – both for the employers themselves but also for the society as a whole. This can be achieved through Corporate Social Responsibility (CSR) initiatives, targeted awareness activities such as in-person and online seminars, as well as concise, well-designed briefing materials that highlight success stories, productivity gains, and the broader societal value of inclusive employment. In this respect, the PES in Greece has established initiatives such as the so-called “Independent–Strong–Free” one, which aims to create employment opportunities for women who have experienced abuse. To increase their negotiating power, employer counsellors who engage directly with employers should be equipped with practical prompts and talking points to be able to effectively communicate the advantages of hiring individuals from disadvantaged groups, dispel misconceptions, and foster more open and informed employer and existing staff attitude. Another way of promoting inclusivity would be to award employers who hire frequently from those groups by providing a “partner” badge or label, as done, for example, by the PES in Luxembourg. This could serve as both a recognition and an incentive, motivating other companies to do the same (OECD, 2025[8]).
Employers’ hesitation may partly stem from limited or inaccurate perceptions of the profiles of individuals facing vulnerabilities. Helping employers better understand these profiles can therefore play an important role in reducing such concerns. In this context, improving the depth and quality of jobseeker profiles – including via the functionalities of the platform – could potentially create more opportunities. Providing employers with more comprehensive information about candidates’ skills and experience can highlight strengths that may otherwise remain overlooked. This may include informal or transferable skills, prior work experience, as well as broader employability attributes such as motivation, adaptability, attitude and communication skills.
Outreach and communication activities play a key role in securing and expanding employment opportunities for individuals facing vulnerabilities, especially if these take place in a proactive manner. Such efforts can help raise awareness about the full range of support services available and the benefits of hiring from these groups. To ensure that information reaches a broad range of employers, outreach efforts should use multiple channels. In addition to personalised contact with companies, PES can rely on wider communication tools such as mass communication methods, printed material, information campaigns, job fairs, or dedicated information sessions. Participation in events organised by other stakeholders can also provide valuable opportunities to present available services and strengthen visibility. Intermediaries such as employer associations, chambers of commerce and professional networks can be particularly effective partners, as they are able to disseminate information to their members and encourage engagement. Most importantly, regular and consistent communication and co‑operation with networks and associations representing individuals with vulnerabilities is crucial. Their inputs can help inform the design and implementation of relevant initiatives while at the same time, their feedback will be key for continuous assessment and refinement. Particular attention should be given to ensuring diverse representation within these networks and associations, reflecting different types of vulnerabilities and experiences. At the same time, these networks serve as core channels through which public agencies can disseminate information and promote initiatives and tools like the matching platform, thereby enhancing greater take up among target populations.
PES could also organise dedicated employer awareness events – either open to all employers or targeted to specific sectors – to provide information on hiring individuals from vulnerable groups and present relevant support measures. These events can be organised in hybrid formats or complemented with virtual participation options in order to maximise accessibility and participation. Similar initiatives already exist in several countries. For example, the “AMS Business Tour” organised by the Austrian PES aims to strengthen employer engagement and address workforce challenges by presenting available services and building closer co‑operation with companies (OECD, 2025[8]).
While it is important to create jobs for these individuals in the first place, equally important is to ensure that these jobs are of good quality and sustainable. Another challenge concerns the quality and sustainability of available jobs. Sometimes, matching platforms optimise volumes i.e. the number of matches or placements, without sufficient attention to long-term outcomes for users. This approach can have problematic implications when one of the key goals is to support vulnerable populations. While increasing access to jobs is of utmost important, placing users into low-quality or unstable roles can reinforce insecurity and uncertainty rather than improve livelihoods and promote self-sufficiency. Therefore, they may find themselves being trapped in dead-end jobs that offer short-term financial relief but ultimately limit future opportunities, for example by leaving little time to pursue training or actively search for better opportunities.
Indeed, the results of the analysis in Chapter 4 highlight that MIS beneficiaries are often concentrated in temporary, part-time, or low-paid employment with limited prospects for advancement. It’s also shown that they frequently return to unemployment after relatively short employment spells (e.g. around eight months). The matching and career recommendation platform can help those individuals identify available vacancies including the ones with better long-term prospects, while also reducing the time between successive employment spells. To be able to do this, the platform should clearly define what constitutes a “good opportunity” and promote it. This includes emphasising key factors such as employment stability (e.g. nature and security of the contract) alongside fair and adequate wages. In addition, it should highlight opportunities that support skill development, offer clear pathways for career progression, and are provided by employers with a strong reputation.
Despite the platform’s capabilities, individuals themselves cannot affect or ensure the quality of those opportunities. Several countries have combined activation policies with measures aimed at promoting more stable employment pathways. For instance, Germany’s integration programmes for long-term unemployed individuals often include structured training and work placements designed to lead to more sustainable employment rather than short-term job placements. Similarly, in Finland, employment services place strong emphasis on training and skills development as part of activation strategies, ensuring that individuals move towards occupations with better long-term prospects. In Slovenia, employers who fail to meet their own obligations or commit labour law violations are recorded in a negative registry and rendered temporarily ineligible for job referrals or financial support (OECD, 2025[8]).
Building on such approaches, policy can further strengthen job quality through concrete measures. This includes setting minimum standards for jobs that are promoted on platforms. Such standards may include – in addition to respect for wage rules –, clear description of tasks and a minimum job duration. In addition, funding from public sources could be tied to real outcomes, like whether people stay in their jobs or move to better positions over time. At the same time, simple systems can be introduced to monitor and signal employer quality. For instance, accreditation or rating schemes can ensure that only employers who meet basic standards are fully included. Obviously, stronger labour inspections and data-sharing mechanisms can help identify patterns of precarious or poor-quality employment.
Policies can also encourage employers to think beyond short-term hiring, for example through co-financing on-the‑job training linked to career progression, rather than just filling immediate vacancies. Finally, a shift in performance measurement is essential. Instead of focussing only on the number of placements, policies should meticulously track what happens afterwards. This includes how long people stay employed, whether their earnings improve, their progression in better roles and other indicators of meaningful and sustainable employment.
ALMPs and other opportunities should be continuously available
Even the most advanced career recommendation tools cannot produce their intended results when there is lack of underlying services. All types of ALMPs – including training first and foremost – but also other types of programmes such as educational opportunities and social services must be available and shown on a continuous basis, so that individuals can see them and act when they are ready, rather than waiting for limited enrolment windows or uncertain programme cycles. This does not require that each and every programme run at all times; yet it means ensuring consistent access through rolling entry points, modular training formats, and baseline services such as job search guidance that are always available.
Naturally, the existence and continuous availability of programmes and services rely on stable and predictable funding. Funding disruptions can create inefficiencies and uncertainty to all. Nevertheless, given the limited resources that PES usually have at their disposal, it is important that funding-related decisions are based on what works well and what is not, which can only happen following rigorous evaluations and assessments. This includes not only programmes per se but also the digital tools themselves (see Section 6.3.4 of this Chapter) for a more detailed discussion on this matter).
Lack of information about available opportunities may lead to discouragement and de‑activation
While ensuring a sufficient number of job opportunities is one fundamental prerequisite for the sound operation of a matching platform, ensuring an adequate and relevant pool of candidates is equally important. Although situations with insufficient candidates may be less common, they do arise under specific occasions. For instance, this can occur in specific sectors (e.g. care work, construction, technical roles) or in certain geographic areas. The issue here is not a lack of labour demand and supply, but rather a mismatch related to skills, attractiveness, or accessibility. Jobs could remain unfilled not because candidates do not exist, but because they are not adequately informed, prepared, or incentivised to take them up.
The platform here can play an instrumental role in reducing information gaps and helping individuals better understand available opportunities and pathways. Through its functionalities and by drawing on multiple data sources, including job vacancy postings, administrative data, occupational classification systems (e.g. ESCO), and labour market information, it can make jobs and training options more visible, relevant, and easier to understand. For example, it can summarise key aspects of a job such as typical tasks, required skills, working conditions and earnings to expect, and if data permit what progression opportunities may follow. It can also guide users towards concrete next steps, such as targeted training or other activation options, helping them come one step closer to employment. If this information is presented in a structured, friendly and personalised way, the platform can make opportunities feel both more realistic and approachable. For instance, by indicating whether a job is accessible within a reasonable commute or compatible with personal constraints.
Although the platform itself can reduce practical frictions including by better filtering, indication of supportive options (e.g. childcare services in nearby radius), actionable suggestions (e.g. this job is reachable in 30 minutes by public transport), it cannot fully address structural barriers. Complementary policy actions are therefore essential. In particular, reducing or eliminating information gaps requires proactive and sustained outreach efforts. Many individuals, especially those in vulnerable situations, may not be aware of existing tools, support measures, or opportunities, or may not fully understand how to access and use them. Without this awareness, the risk of discouragement and disengagement is very high, especially for those further from the labour market. It is therefore essential for policymakers to invest not only in advanced digital solutions, but also in outreach strategies and relevant activities. Such activities may include career guidance and awareness campaigns that actively reach target groups, raise awareness of available support, and clearly explain how digital tools can be used and benefit them in practice. Ensuring that people understand both the opportunities available to them and how to navigate digital platforms is key to ensure increased participation and continuous engagement.
In order to match jobseekers to vacancies effectively and provide effective career advice, digital job matching and career recommendation tools need to draw on multiple types of data held by different institutions in both Greece and Belgium. This, in turn, calls for an IT infrastructure that supports secure, real-time data exchange and linkage across systems and organisations. Moreover, the quality and accuracy of the recommendations produced by such tools critically hinge on the coverage and quality of the underlying data. Limited coverage restricts the tools’ usefulness and take‑up, while poor-quality data undermine the relevance and robustness of the recommendations.
This section provides recommendations on the data that could most effectively support job matching and career recommendation tools for Greece and Belgium, as well as measures to improve data coverage and quality. It also briefly outlines some key legal considerations governing the use of these data for each country separately.
6.4.1. In Greece, five types of data can support the matching tool, but an expanded infrastructure is required for real-time linkages and exchanges
Based on the analysis, the key recommendations for Greece include:
The job matching tool should draw on information on skills, competences and employability from the GMI platform and DYPA, complemented by employment histories from ERGANI.
Data on demographic characteristics should be used for monitoring and evaluation – in particular, to assess patterns of use of the tool and potential biases – but not as direct inputs in the matching process.
The web services provided by KED offer a robust infrastructure that could be further expanded to ensure that all data necessary for the well-functioning of the tool are exchanged and linked securely and timely.
Jobseekers and employers in registering occupational information, and using a common occupational nomenclature across registries, can improve data quality and reduce the risk of coding errors.
Constructing jobseeker employment histories by drawing not only on ERGANI but also on social security data can improve coverage by incorporating information on self-employment (and other non-salaried work) and employment in the public sector. At the same time, survey-based measures of skills and competences (such as DYPA’s profiling questionnaire) remain essential to capture skills acquired through informal employment and other forms of work.
In general, decisions on data architecture and infrastructure for job matching tools need to be closely aligned with the overall governance of the tool. Two strategic governance decisions will have important implications for the underlying data architecture of the job matching tool. The first strategic decision concerns institutional ownership. This choice will determine which data exchange agreements need to be established, between which institutions and on what technical basis. DYPA already holds extensive labour market data in its information system, which is well suited to support a job matching tool, as these data are collected explicitly to underpin DYPA’s processes and activities aimed at reintegrating jobseekers into the labour market. Moreover, DYPA already operates a job platform through which jobseeker can browse vacancies and employers can search for candidates, and is in parallel developing an advanced competency-based job matching tool. As these tools are targeted to all jobseekers registered with DYPA, they are also intended to serve MIS beneficiaries registered with DYPA. There is therefore considerable scope to exploit synergies and avoid duplication. At the same time, primary responsibility for GMI beneficiaries lies with the Ministry of Social Cohesion and Family.
The second, closely related strategic decision concerns the target group of the tool. If the tool is intended to serve only MIS beneficiaries under the third pillar (i.e. those with an obligation to register with DYPA), DYPA’s IIS could provide the primary data backbone for job matching. DYPA data are already collected with the explicit purpose of supporting the labour market integration of registered jobseekers and are already now used for job matching. However, restricting the tool to DYPA data alone would exclude a non-negligible share of GMI beneficiaries who are not required to register with DYPA, for example because they are employed at the time of applying for the scheme. If the target group of the tool is to include these GMI beneficiaries as well, the variables used for job matching will need to be collected through other channels – for instance, at the point of registration on the MIS platform, or by extending the obligation to register with DYPA to MIS beneficiaries who declare to be employed at registration for GMI but whose household income falls below the MIS eligibility threshold.
Regardless of these governance choices, the web services provided by the Interoperability Center (KED) of the Ministry of Digital Governance offer a robust infrastructure for data exchange, and their use could be scaled up beyond current applications. At present, basic variables – such as gender, age and education – are collected separately by multiple institutions, often using different definitions or formats. This can lead to inconsistencies across registries and undermine the reliability of administrative data. Designating a single institution as responsible for collecting and updating these core variables and ensuring that other institutions access them as needed web services, would both reduce errors and significantly improve data quality, while also support compliance with the GDPR “once‑only” principle of data collection, by limiting the need for multiple bodies to repeatedly collect and store the same personal information. This principle is increasingly common across OECD countries, including for example Belgium, Austria and the Netherlands (OECD, 2023[9]) In doing so, it is crucial to ensure that all data exchanges are lawful by establishing appropriate data exchange agreements between the institution hosting the data underpinning the tool and the institutions that own the source data (Box 6.1).
Regarding the allocation of responsibilities under the GDPR, DYPA should assume the role of the Data Controller for the new Job Matching Tool. This is consistent with DYPA’s statutory mandate to provide labour market services and prevents the administrative complexity and diluted accountability associated with the alternative of joint controllership arrangements.
The initial collection of administrative data in the context of the operation of other databases is lawfully grounded in the performance of tasks of public interest. Yet, a systematic reuse and linkage of these datasets for an AI-enabled tool presents legal uncertainties under current frameworks. In particular:
Reliance on the “further processing” compatibility mechanisms of Article 6(4) GDPR is insufficient.
Existing national provisions in Law 4 624/2019 do not encompass the systematic reuse of data in the context of operation of the Job Matching Tool.
Consequently, it is strongly advised that a specific legislative provision be adopted to provide a clear legal basis, define administrative procedures and specify technical and organisational safeguards, in light of Articles 6(1)(e) and 9(2)(g) GDPR.
As the Job Matching Tool retains human oversight, it does not constitute “solely” automated decision making under Article 22 GDPR. However, transparency obligations remain appliable, requiring that beneficiaries be informed of the tool’s role and provided with meaningful explanations regarding its outputs.
It is not yet possible to determine with certainty the extent to which the Job Matching Tool will make use of AI techniques within the meaning of the EU AI Act. The final qualification of the tool will depend on the concrete design and implementation choices to be made by the Greek authorities after the end of this project. To the extent that the Job Matching Tool fulfils the relevant conditions laid down in the EU AI Act, further compliance obligations would arise from the regulation.
More specifically, the Job Matching Tool has a potential to be classified to be a high-risk tool according to the EU AI Act and the European Data Governance Act (DGA). The Job Matching Tool does not assess eligibility for social benefits (which would trigger high-risk status under Annex III, 5a) in its current high-level design proposed in the project. However,), it may be classified as high-risk under Annex III, 4a if its profiling functionalities exert a substantive influence on recruitment and employment outcomes.
If the new tool will be classified as high-risk, the competent public authority acting as deployer would be required to comply with the applicable obligations under the EU AI Act and the corresponding national implementation framework established by Law 5 321/2026. These include, where applicable, conducting a Fundamental Rights Impact Assessment prior to deployment, ensuring effective human oversight and appropriate quality of input data, following the provider’s instructions for use and maintaining system logs.
6.4.2. In Belgium, five data categories can support a digital job matching tool, but stronger interoperability and more consistent coding are needed for efficient, near real-time data exchange
Based on the analysis, the key recommendations for Belgium include:
The job matching tool should integrate information from both CPAS/OCMW and regional PES systems, combining CPAS data on education, household circumstances and integration pathways with PES profiling data on occupations, skills and work preferences, and with employment histories from ONSS/RSZ and Sigedis.
Demographic variables (such as age, gender, nationality and household composition) should be used for monitoring and evaluation rather than for matching, in order to detect potential biases in algorithmic recommendations and ensure fairness and transparency.
Belgium’s CBSS interoperability infrastructure could be further used to support co‑ordination among institutions, helping CPAS/OCMW, PES, ONSS/RSZ and Sigedis exchange data in secure, timely and standardised ways. Strengthening this co‑operation would help reduce duplication and promote greater consistency in core variables across registries.
Federal actors could encourage closer alignment around common occupational and skills taxonomies (such as ESCO/ISCO) across CPAS/OCMW and PES systems. Supporting such harmonisation efforts would enhance coherence in how skills and occupations are recorded, reduce coding inconsistencies, and make it easier to match competencies to vacancies across regions.
Information on employment barriers could be made more comparable across institutions, for example by promoting the use of shared tick-box categories in CPAS and PES intake forms (e.g. childcare constraints, transport availability, health limitations). Clarifying and co‑ordinating these definitions through a common framework would improve the consistency of these variables and their usefulness for job matching tools.
Belgium’s administrative data environment offers strong foundations for developing a job matching tool for MIS beneficiaries, but its architecture and governance require several strategic choices. Responsibility for beneficiary data is shared across institutions. CPAS/OCMW centres, which administer social assistance, hold detailed case information, often stored in Nova PRIMA/Nova+. These records cover education, qualifications, household circumstances and participation in integration pathways. Regional PES (VDAB, Le Forem, Actiris and ADG) collect more systematic labour market information when individuals register as jobseekers, including their work history, competencies and preferred occupations.
A key governance decision concerns institutional ownership of the tool. If the intended beneficiaries are social assistance recipients who are already registered with a regional PES, regional PES databases could form the primary backbone, given that these data are collected specifically to support labour market integration pathways. However, not all MIS beneficiaries are required to register with a PES, meaning that relying exclusively on PES data may exclude part of the target group. In that case, CPAS/OCMW systems would need to provide the additional variables required for matching, such as occupations, skills and work preferences.
Another strategic question concerns how information from different systems should be integrated. Belgium already has robust interoperability through the CBSS, which connects the federal population register, ONSS/RSZ employment records (via DIMONA and DmfA), Sigedis career data and CPAS/OCMW case files. CPAS/OCMW can already retrieve verified career histories and benefit information through CBSS services. However, basic variables such as education, skills and availability are still collected separately by CPAS/OCMW and each PES, often using different formats. Assigning clear responsibility for maintaining core beneficiary variables and ensuring they are reused by other bodies through CBSS exchanges rather than recollected, would reduce duplication and improve consistency.
Aligning data architecture with governance decisions is therefore essential. If Belgium aims for a tool that covers all RIS beneficiaries, then structured data would need to flow reliably from CPAS to PES systems and vice versa, using common definitions embedded in Nova PRIMA and PES profiling tools. If instead the tool focusses only on beneficiaries engaged with PES services, then PES systems could remain the main data hub, with CPAS providing only selected updates. In both cases, clear data‑exchange agreements for sharing are required to ensure compliance with GDPR and to uphold the principles of data minimisation and proportionality (Box 6.2).
The development and use of an AI-powered job matching tool to support the matching of MIS beneficiaries with labour demand should take into consideration key legal considerations under Belgian and EU data protection laws. These key considerations include: 1) the role of the institutions under the GDPR; 2) the legal basis for sharing and using personal data to develop and use the Tool; 3) the need for a data protection impact assessment; 4) the use of subcontractors; 5) the potential application of automated decision making.
Regarding the role of institutions under the GDPR, each institution involved in the development and use of the job matching tool must assess its role under the GDPR to ensure GDPR compliance. Institutions that directly influence the purpose or means of the job matching tool act as a controller. These institutions should determine if they qualify as joint controllers (they determine the purpose or means jointly with other controllers) or separate controllers (they determine the purpose or means separately from other controllers). In case of joint controllership, a joint controller agreement is required between all joint controllers. In case of separate controllership, requirements vary depending on whether the involved institutions sharing personal data qualify as public authorities or social security institutions. If they qualify as public authorities, a protocol between the separate controllers is required. However, if they qualify as social security institutions, a prior deliberation by the social security and health chamber of the Information Security Committee – instead of a protocol – is required.
Unlike controllers, institutions that solely link and provide data without directly influencing the purpose or means of the job matching tool act as a processor. In such cases, a data processing agreement is required. Considering the planned set-up of the job matching tool, however, no institution seems to qualify as a processor. Importantly, a comprehensive assessment of the role of each institution would require knowledge of practical aspects of the implementation of the job matching tool and must therefore be conducted once these practical aspects have been decided.
With respect to the legal basis for sharing and using personal data to develop and use the job matching tool, each involved institution already undertakes some data processing to perform its current activities (i.e. the initial processing activities). In addition to these activities, each institution will also undertake additional processing activities in the context of the Tool. For the purpose of this assessment, it is useful to distinguish between the additional activities involved in the development and those involved in the use of the job matching tool. Under the GDPR, each of these processing activities must rely on a legal basis. To determine the legal basis required, it must first be determined whether the processing activities in the context of the Tool have a different purpose than the initial processing activities. If the purpose is the same, then the processing activities in the context of the job matching tool can rely on the same legal basis as the initial processing activities. Based on a preliminary assessment, in Belgium this would be the case for the use of the job matching tool by certain institutions. If the purpose is different, a compatibility test is required to determine whether the processing activities in the context of the job matching tool are compatible with the initial processing purpose. Based on a preliminary assessment, in Belgium this would be the case for the use of the job matching tool by certain institutions, and for the development of the job matching tool by all institutions. In case the compatibility test succeeds, the processing activities in the context of the job matching tool can rely on the same legal basis as the initial processing activities. Otherwise, the processing activities in the context of the job matching tool must rely on a separate legal basis. In this case, it is likely that a legislative intervention amending the relevant laws will be required to enable the institutions concerned to undertake the processing activities in the context of the job matching tool. A comprehensive assessment of the legal basis or the performance of the compatibility tests would require knowledge of practical aspects of the implementation of the job matching tool and must therefore be conducted once these practical aspects have been decided.
Each institution acting as a controller must perform a data protection impact assessment for both the development and use of the job matching tool. If subcontractors (i.e. service providers) will be involved in developing and maintaining the job matching tool’s digital infrastructure, they will process personal data on behalf of the institution(s) and therefore qualify as processors under the GDPR. To ensure compliance with Belgian law, the following considerations must be taken into account. First, a data processing agreement must be concluded between each processor and the parties that engage or designate them (likely to be one or more of the leading institutions). Second, the service providers must appoint a data protection officer. Third, public procurement procedures must be followed.
Regarding the potential application of automated decision making, if the job matching tool enables decision making by technological means without human involvement (i.e. automated decision making) or enables the use of personal data to evaluate certain personal aspects relating to a natural person – particularly to analyse or predict aspects concerning that natural person’s performance at work or economic situation (i.e. profiling) – then explicit consent from each data subject for whom the job matching tool provides job-matching advice or recommendations may be required, or the national legal framework may need to be amended to allow for the processing. To ensure the job matching tool’s practical feasibility, its design and operation should be structured so as not to constitute automated decision making or profiling. Practical measures may include:
Ensuring that the job matching tool interface provides a warning that the output should merely complement rather than replace professional judgement, should be double‑checked by the user and should never be trusted as the sole basis for taking action,
Ensuring that the output can always be overridden by the user,
Educating and training users of the job matching tool to prevent users to rely heavily on the output delivered (e.g. regular trainings, access to original data sources for manual recommendations), and
Adopting a job matching tool policy prohibiting relying solely on the output delivered by the job matching tool for taking action.
6.4.3. In Greece, strong existing data can support the career recommendation tool, though further improvements in data quality and coverage are needed
As outlined in Chapter 5, the career recommendation tool is intended to form an additional pillar complementing the job matching tool (or a standalone career recommendation tool). Based on this analysis, the key recommendations for Greece include:
Employment data from ERGANI could serve as the backbone of the data underpinning the career recommendation tool. The quality of data on hirings, contract changes and contract ends could be improved through further digitalisation of data collection. Ongoing efforts, including the transition to ERGANI II, offer important opportunities in this regard.
M.E.K.Y. already uses ERGANI and job vacancy data within its Diagnostic Mechanism to monitor labour market developments and generate indicators of labour demand across occupations and industries. These indicators could be integrated into the career recommendation tool.
Enhancing DYPA data on ALMP participation by including key programme parameters would help provide a more comprehensive picture of individuals’ career trajectories.
Data in DYPA’s IIS and Diofantos could be used to map available ALMPs and feed directly into the tool to recommend suitable ALMPs to MIS beneficiaries.
Greece could consider developing a comprehensive mapping of relevant social services to better support MIS beneficiaries, which could be integrated into the tool to direct users towards social services that are complementary to employment services.
The same strategic considerations regarding institutional ownership of the job matching tool also apply to the career recommendation tool. In particular, the choice of institutional ownership will determine which data exchange agreements are required and what data sharing services need to be put in place or expanded. Given that the backbone of the data underpinning the career recommendation tool consists of ERGANI data augmented with information on ALMP participation held by DYPA and Diofantos, DYPA already owns most of the data required and could expand its existing web services with ERGANI to retrieve any additional information needed. In this respect, it is important to ensure that all data exchanges are lawful by establishing appropriate data exchange agreements between the institution hosting the data underpinning the tool and the institutions that own the source data (Box 6.3).
The career recommendation tool is conceived as part of a broader, integrated and AI-enabled job matching platform, developed as an additional pillar extending the functionality of the envisaged job matching tool. At the same time, an alternative design scenario is envisaged, under which the tool would be developed as a separate, standalone, digital instrument. Notwithstanding differences in technical configuration, both scenarios envisage interoperability with existing systems and the use of data drawn from existing databases, including, as appropriate, data from the MIS (GMI) platform, DYPA systems, ERGANI and Diofantos.
A key requirement under the GDPR is to clearly define the roles and responsibilities of the different actors involved in processing personal data. These roles determine who is responsible for ensuring compliance with data protection rules and depend on how each actor uses and controls the data. In this context, the present assessment takes the view that, irrespective of which of the two envisaged design scenarios is ultimately adopted, DYPA should assume the role of controller for the operation of the career recommendation tool. This applies both where the tool is developed as part of a broader integrated platform and where it is developed as a separate, but interoperable, digital instrument. Such allocation is consistent with DYPA’s statutory mandate in the field of activation, counselling, employability support, labour market integration and training-related measures. It also avoids the administrative complexity and dilution of accountability often associated with joint controllership arrangements. The other authorities involved, including the Ministry of Social Cohesion and Family and the Ministry of Labour and Social Security, would in principle remain independent controllers in relation to their respective systems and processing activities.
Beyond the allocation of roles, the use of personal data by the tool raises several important legal questions under the GDPR, including the legal basis for processing, the reuse of existing data, data sharing between authorities and the safeguards required when linking different datasets. In simple terms, the GDPR allows personal data that were originally collected for one purpose to be reused for a different purpose only under certain conditions. These conditions require, among other things, that the new use is compatible with the original purpose for which the data were collected and that appropriate safeguards are in place. In the case of the career recommendation tool, the intended use of data – namely combining and analysing information from multiple public databases for the enabling of all the tools functionalities – goes beyond what these existing rules provide for. In particular, the systematic reuse and linkage of data across different systems for this new purpose cannot be safely based on the existing “further processing” mechanism alone, while existing national provisions do not encompass the systematic reuse of data in the context of operation of the career recommendation tool. For this reason, it is strongly advised that a specific legislative provision be adopted to provide a clear legal basis, define administrative procedures and specify technical and organisational safeguards, also taking into account any legislative provision designed for the operation of the job matching tool.
With regard to automated decision making, the career recommendation tool, as currently envisaged, does not constitute a system of solely automated decision making, since its outputs are intended to remain recommendatory and final decisions relating to participation in Active Labour Market Policies (ALMPs) and referrals to other services remain subject to human assessment.
As for the EU AI Act (the EU’s regulatory framework governing artificial intelligence systems), it is not yet possible to determine with certainty the extent to which the career recommendation tool will make use of AI techniques within the meaning of the EU AI Act. The final qualification of the tool will depend on concrete design and implementation choices. To the extent that the career recommendation tool fulfils the relevant conditions laid down in the AI Act, further compliance obligations would arise from the regulation.4
System classification – High-risk potential: while the career recommendation tool does not, in its currently envisaged design, appear to assess eligibility for the MIS scheme or other essential public assistance benefits and services, and therefore does not at this stage appear to fall within Annex III, point 5(a), certain of its functionalities may potentially fall within Annex III, point 4(a), to the extent that they involve profiling and materially influence access to employment, training, activation measures or related opportunities. In addition, certain functionalities of the tool may also raise considerations under Annex III, point 3(a) (education and vocational training), if and to the extent that its outputs are used to determine or significantly influence access to, or assignment to, educational programmes;
Compliance obligations: if classified as high-risk, the competent public authority acting as deployer would be required to comply with specific obligations, including the establishment of effective human oversight, the use of high-quality input data, adherence (if applicable) to the provider’s instructions for use, and the maintenance of system logs.
6.4.4. Belgium has a strong data foundation for a career recommendation tool, but greater interoperability and consistency are needed to overcome fragmentation
Given that most of the data requirements are common to both the career recommendation tool and the job matching tool, the recommendations for job matching also largely apply to the career recommendation. Based on the analysis, the key recommendations for Belgium include:
Federal social security data (ONSS/RSZ and Sigedis) can form the backbone of the career recommendation tool, enabling the construction of detailed employment trajectories and occupational transitions relevant for MIS beneficiaries.
Aggregated labour market intelligence produced by regional PES (VDAB, Le Forem, Actiris and ADG) – such as shortage occupation lists and sectoral trends – should be integrated into the tool to ensure recommended pathways are aligned with sustainable labour market demand.
Linking employment histories with data on ALMP participation, including training and public works programmes (e.g. Article 60), is essential to reflect the actual career pathways of MIS beneficiaries and improve the relevance of recommendations.
Up‑to‑date information on available ALMPs, particularly training programmes, should be systematically mapped and integrated, allowing the tool to recommend combined “occupation + training” pathways based on identified skill gaps.
Greater standardisation and structuring of skills and employability data collected by CPAS/OCMW, and better alignment with PES profiling systems, would strengthen personalisation and facilitate data reuse through the CBSS infrastructure.
Belgium’s CBSS interoperability framework provides a strong foundation for data exchange, but clearer governance and wider use of common classifications for skills and barriers would further enhance data quality for career recommendation purposes.
The choice of institutional ownership and governance arrangements for the career recommendation tool will be closely linked to those for the job‑matching tool. If the tool is intended to cover all MIS beneficiaries, including those not registered with a PES, CPAS/OCMW systems would need to play a central role alongside regional PES databases.
Belgium’s CBSS infrastructure already provides secure, real‑time data exchange between CPAS, PES and social‑security institutions. However, differences in how similar variables (e.g. skills, employability, constraints) are recorded across institutions limit the effective reuse of data. Clarifying responsibilities for maintaining core variables and promoting systematic reuse through CBSS exchanges – rather than repeated collection – would improve consistency and data quality, in line with the “once‑only” principle widely promoted across OECD countries. In this respect, it is important to ensure that all respective data exchanges are lawful (Box 6.4).
As the career recommendation tool is structured in a similar way to the job matching Tool, most elements discussed for this tool (see Box 6.2) also apply to the career recommendation tool.
1. Findings that apply to both tools
Both tools have similar governance arrangements, so the roles of the participating organisations under the GDPR are also similar. In addition, several other considerations apply to both tools, including: whether a Data Protection Impact Assessment (DPIA) is required under the GDPR, whether and how the EU AI Act applies, including, where the tools qualify as high-risk AI systems, requirements concerning Fundamental Rights Impact Assessments and other obligations arising from this classification; how to manage outside contractors who help build or run the tool; whether the AI Act (a new law regulating artificial intelligence) applies; and the possible effects of the EU’s simplification proposal on digital laws.
Because both tools are designed to help people find work or access social support, similar considerations apply regarding the legal grounds for using personal data and whether a “compatibility test” is needed. A compatibility test checks whether using data for a new purpose is compatible with the original purpose the data was collected for. Data subject profiling, i.e. building a profile of a person based on their data, will likely be a core feature of both tools.
2. Findings specific to the career recommendation tool
Unlike the job matching tool, the career recommendation tool is unlikely to affect MIS beneficiaries’ access to a job. This is because the career recommendation tool only provides non-binding suggestions about possible career paths, training courses, or social and psychological support services, it does not make decisions or create an output that directly control or influence whether a MIS beneficiary gets a job.
Because the career recommendation tool may be used directly by individuals (such as MIS beneficiaries) without professional guidance, extra safeguards should be built in. For example, users should have clear contact details for staff at the relevant organisations if they have questions about the career recommendation tool’s output or how it works. There should also be plain guidance explaining that the career recommendation tool’s suggestions may not always be accurate or complete.
References
[1] Brioscú, A. et al. (2024), “A new dawn for public employment services: Service delivery in the age of artificial intelligence”, OECD Artificial Intelligence Papers, No. 19, OECD Publishing, Paris, https://doi.org/10.1787/5dc3eb8e-en.
[3] Council of the European Union and European Parliament (2016), Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC, (General Data Protection Regulation) (Text with EEA relevance), http://data.europa.eu/eli/reg/2016/679/oj.
[8] OECD (2025), Optimising Processes and Services at Bulgaria’s National Employment Agency, Connecting People with Jobs, OECD Publishing, Paris, https://doi.org/10.1787/4e79e9db-en.
[2] OECD (2024), Recommendation of the Council on Artificial Intelligence, OECD/LEGAL/0449, https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449 (accessed on 18 September 2025).
[7] OECD (2023), Impact Evaluation of the Digital Tool for Employment Counsellors in Spain: SEND@: Report on the Design and Implementation of an Impact Evaluation of the Digital Counselling Tool for Spain’s Public Employment Services, OECD Publishing, Paris, https://doi.org/10.1787/fe1ec3c3-en.
[9] OECD (2023), Personalised Public Services for People in Vulnerable Situations in Lithuania: Towards a More Integrated Approach, OECD Publishing, Paris, https://doi.org/10.1787/e028d183-en.
[6] OECD (2022), Harnessing digitalisation in Public Employment Services to connect people with jobs, OECD Publishing, Paris, https://doi.org/10.1787/367a51f3-en.
[4] OECD (2021), “Building inclusive labour markets: Active labour market policies for the most vulnerable groups”, OECD Policy Responses to Coronavirus (COVID-19), OECD Publishing, Paris, https://doi.org/10.1787/607662d9-en.
[10] OECD (n.d.), OECD AI Principles overview, https://oecd.ai/en/ai-principles (accessed on 27 November 2025).
[5] UWV (n.d.), Onze algoritmes [Our algorithms], https://www.uwv.nl/nl/over-uwv/organisatie/algoritmeregister-uwv/onze-algoritmes (accessed on 14 May 2024).
← 1. The OECD AI Principles were initially adopted by Member countries in May 2019 as part of the OECD Council Recommendation on Artificial Intelligence and updated in May 2024. The Recommendation sets out five values-based principles for the responsible stewardship of trustworthy AI, alongside related recommendations for governments to include in their national policies and international co‑operation in this domain (OECD, n.d.[10]).
← 2. SMS (Short Message Service): interaction through sending and receiving text messages; USSD (Unstructured Supplementary Service Data): a simple text-based menu used on basic mobile phones (e.g. dialling short codes like *123#); IVR (Interactive Voice Response): a phone‑based system where users navigate menus or receive information through automated voice prompts.
← 3. An extensive set of Key Performance Indicators (KPIs) for the AI-powered matching and career recommendation platform has been developed as part of the project and can be accessed on the project’s webpage.
← 4. As the tool’s technical specifications are further developed and its functionalities evolve, it may be appropriate to update the analysis to reflect its final classification as an AI system.