This chapter explores how artificial intelligence can address persistent challenges in social security across the EU. Drawing on examples from Catalonia, Germany, and Finland, it highlights how AI tools are being used to support eligibility checks, categorise unstructured data, and automate document processing. These cases demonstrate the potential of AI to improve efficiency, reduce burdens on both staff and citizens, and enhance access to benefits and services. However, they also reveal the need for critical enabling conditions: solid digital infrastructure, data interoperability, and clear governance frameworks.
Harnessing Artificial Intelligence in Social Security
Use Cases, Governance and Workforce Readiness
Abstract
Social security institutions across OECD countries face operational challenges that limit their ability to provide social protection: fragmented data management, low interoperability, limited outreach to vulnerable eligible populations, and insufficient capacity in terms of workforce and skills (OECD, 2024[1]). These difficulties are exacerbated by broad trends such as demographic shifts, economic constraints, and environmental changes (OECD, 2024[2]).
In this context, artificial intelligence (AI) has emerged as a potential tool to support public administrations in the design and delivery of services. When thoughtfully implemented within a broader strategy, AI systems can help automate tasks, streamline processes, reduce administrative burden, and improve service delivery (OECD, 2025[3]; OECD, 2025[4]).
At the same time, the use of AI in the public sector comes with potential risks and requires a robust governance framework to ensure trustworthy implementation, especially in the field of social security (OECD, 2025[4]). These risks include the mishandling of personal data, lack of transparency in automated decision-making, algorithmic bias, and the potential to reinforce existing structural inequalities. Addressing these challenges requires building the right enablers for trustworthy AI adoption, implementing guardrails to manage risks, and engaging stakeholders to incorporate a broad range of perspectives. These governance mechanisms will be examined in detail in Chapter 2.
Methodology and use cases
This chapter explores how AI can be leveraged to address challenges in social protection, through three use cases from public institutions across Europe. The analysis is based on desk research, internal documentation provided by the institutions, and direct input from institutional stakeholders, collected via interviews and informal discussions.
These interviews have been guided by the pillars of the OECD Framework for Trustworthy Artificial Intelligence in Government (OECD, 2025[3]), which provided a structure for assessing the use cases. This same framework will be applied in greater detail in Chapter 2.
Use cases have been selected based on the following criteria:
Institutional and geographical diversity, with a focus on EU Member States (wherever possible) given this report is funded by the EU via the Technical Support Instrument.
Maturity of the implementation, prioritising AI systems already in production for which impact could be assessed.
Replicability, ensuring that solutions can easily be implemented by other public sector organisations.
Trustworthiness, selecting AI implementations with guardrails embedded in governance processes.
Problem-solution fit, favouring cases where AI was selected as the best technology to address a challenge, not as a default or experimental choice.
The limited number of mature and well-documented AI use cases in the field of social security constrained the selection. However, the chosen examples reflect diverse institutional contexts that can provide valuable insights for other administrations.
Dimensions of analysis
The analysis of use cases in this chapter draws from the OECD Framework for Trustworthy Artificial Intelligence in Government (OECD, 2025[3]) and the OECD Framework for the Classification of AI Systems (OECD, 2022[5]). Each use case is assessed along five dimensions:
Problem definition and institutional gaps: What challenge is the AI system aiming to solve? What gaps in terms of capabilities or service delivery are being addressed?
AI system description: What type of system is used? What is its function (e.g. classification, prediction, automation)? How does it qualify as an AI system under the OECD definition?
Implementation conditions: What political, legislative, technical, and organisational factors enabled the deployment of the solution?
Outcomes and key lessons: What results were achieved? What were the organisational outcomes?
Challenges and limitations: What are the risks of the AI solution as it has been implemented? What have been the challenges in the implementation of AI?
Transferability across administrations: What lessons can be drawn for other social security institutions, or other public administrations? What are the preconditions for the solution to be applicable to another context?
The use cases presented in this chapter also inform the analysis in the next chapters. In particular, AI governance frameworks are explored in Chapter 2, and workforce implications in Chapter 3.
AI holds significant potential to improve the coverage, effectiveness, and efficiency of social protection. The potential application areas are vast: in addition to reducing non-take up of social programmes, AI could help to improve policy design, claims administration, service delivery and monitoring and evaluation (OECD, 2025[4]) (see Figure 1.1).
Schematic of current and likely upcoming use cases for AI and social protection
Note: OECD overview of current and hypothetical uses of AI for social protection.
Source: (OECD, 2025[4]).
AI is already being deployed to improve client support, automate back-office processes and support fraud detection. However, more advanced uses of AI – for example in predictive analytics – are not yet commonplace and routine use of AI to reduce non-take-up is not yet common across OECD countries (OECD, 2025[4]) (see Box 1.1).
The OECD’s assessment of current and potential uses of AI in social protection systems across OECD countries has shown that present-day uses of AI in social programmes are focussed largely on client support, automation of back-office processes, and fraud detection. Several countries have already deployed AI-powered chatbots to extend information and support to citizens. Countries such as Austria, Canada, Finland and the United States have used AI to automate and streamline back-office processes, for example by processing and classifying documents, distributing and directing incoming mail and phone calls, and/or prioritising claims processing by predicting where claims re-calculations are likely to impact claimants.
While not yet commonplace across OECD countries, some of the most potentially rewarding uses of AI in social protection are in the domain of predictive analytics, including to improve outreach and reduce non-take-up in social programmes, as well as to forecast demand, support prevention and early identification and tailor interventions to the needs of clients. Examples of these types of AI uses exist, with researchers and governments having piloted (and to a lesser extent implemented) AI tools to predict the risks of adverse events such as homelessness, domestic-violence recidivism and workplace safety accidents, facilitating both prevention and early intervention.
Governments are proceeding with caution as they identify policy challenges that AI is well-suited to address, establishing guardrails and piloting small-scale projects to ensure that AI applications can be rolled out safely at scale. OECD research shows that this caution is mirrored by the public, where support for AI use is tepid: across the 27 countries participating in the OECD’s Risks that Matter Survey, less than half of respondents (40%) feel that the use of AI to help process and approve social programme applications is good for users. Greater public engagement and strong governance are needed to build trust in the responsible use of AI for social programmes.
Source: (OECD, 2025[4]).
In the Spanish region of Catalonia, AI has been successfully applied to help address energy poverty by automating the generation of eligibility reports for means-tested benefits. This use case highlights the role of a central digital infrastructure supplier in co-ordinating social security benefit providers across the region, improving data sharing, and helping reduce non-take-up through automation.
Problem definition and institutional gaps
Energy poverty is a pressing issue in Catalonia, exacerbated by rising energy costs. A 2015 regional law requires energy providers to maintain a minimum level of supply for insolvent customers who meet specific eligibility criteria, such as low income or vulnerable living conditions (Catalunya, 2015[6]). The implementation of this measure faced a common challenge for administering means-tested social security: the difficulty of identifying eligible households, as data on income, assets, or household composition is often dispersed across different systems and not always shared between authorities (Pye et al., 2017[7]).
The procedure to assess households’ eligibility for continued supply and to eventually grant this measure involves the exchange of data between multiple actors, and actions by the household to grant consent for data exchange and benefit acceptance. When a household stops paying its bills, the energy provider notifies the municipal social services, which then request additional consent, if required, to collect and process socioeconomic data.
After manual verification, if the household is deemed at risk of social exclusion (i.e. having support ended), the municipality informs the provider that they must maintain basic energy supply. This process includes five data exchanges, three consent steps, and a manual assessment, contributing to high non-take-up rates.
Municipalities, as social security benefits providers, are responsible for determining whether households meet the eligibility criteria to have energy provision maintained, but the necessary data are fragmented across multiple administrations and the verification process is complex and labour-intensive. This complexity, combined with strict privacy requirements for sharing information with energy providers, creates significant administrative burden.
Although the 2015 law requires the introduction of a protocol for data sharing, implementation has been facing technical and legal barriers, including the requirement for explicit consent under data protection and social security laws, which applies to households in arrears on energy bills. These constraints have contributed to high rates of non-take-up by potential beneficiaries of the measure. The AI system under study intervenes precisely at the eligibility assessment stage, automating data collection and verification to reduce administrative burden and ultimately improve access to the benefit.
AI system description
To reduce the administrative complexity of the eligibility assessment procedure, the Open Administration Consortium of Catalonia (AOC), a public consortium providing digital public infrastructure in the region, developed an AI system that automates the generation of energy poverty reports. The system is offered as a Software-as-a-Service (SaaS) to municipalities and is designed to support social workers in verifying eligibility for the basic energy supply benefit introduced in the 2015 law (AOC Consortium, 2023[8]).
The system operates on three principal platforms: Hestia, which serves as the local social services information system; Via Oberta, responsible for secure data exchange among national, regional, and local entities via an interoperability catalogue; and a business process management (BPM) solution that streamlines and automates operations. Every month, the energy provider identifies customers experiencing insolvency and transmits the pertinent information to the municipality. Civil servants then upload the data into the AOC platform, which retrieves the necessary socioeconomic parameters for the household, verifies eligibility, and generates a pre-filled report in PDF format, to be sent back to the energy company.
The key component is a deterministic AI model, which follows a predefined set of rules to assess eligibility. Unlike probabilistic models, which provide uncertain outputs through inference, a deterministic model produces the same result every time when given the same inputs. This ensures that the decision logic is transparent and reproducible, allowing people affected by the system to understand and challenge the outcome if necessary (OECD, 2022[5]). The system qualifies as AI under the OECD definition because it is machine-based, operates with an explicit objective (eligibility verification), and infers outputs (eligibility reports) from multiple data inputs to influence decisions in a social protection context.
By automating data collection, verification, calculations, and report generation, the AI system has taken over tasks that were previously performed manually by municipal staff, often involving repetitive checks across multiple databases and manual report drafting. This significantly reduces administrative workload and the risk of human error. While full automation is not yet possible due to legal constraints related to the consent of beneficiaries, the system represents a first step towards proactive delivery of benefits to the entire eligible population.
Implementation conditions
The successful deployment of the AI system was made possible by a combination of enablers, including the legal framework, institutional context, and existing technical infrastructure.
The 2015 regional law established a legal basis for data sharing between energy providers and public administrations, creating the use case for the AI system. However, the EU’s General Data Protection Regulation (GDPR) and the Catalan law on social services require consent from households at several points. First, the municipality requires authorisation to generate the energy poverty report, as it is considered an application for a social benefit. Next, no objection (implied consent) is needed to verify the applicant’s social and economic situation by accessing external data. Finally, explicit consent from the other family members is required to access their income data from the Tax Agency and other health sensitive information, in accordance with the GDPR. These requirements prevent full automation and keep users involved in the process (Catalunya, 2017[9]).
The institutional context and the co-ordination of stakeholders have been major enablers, also allowing for shared IT supply. The AOC, acting as digital infrastructure provider for the region, was able to bring together centralised investment, and the right digital skills and talent. The platforms provided free of charge by AOC (Via Oberta, Hestia, and business process management) allowed local authorities to use secure data sharing systems, lowering maintenance and support costs.
Outcomes and key lessons
The main outcome of the AI system for automating energy poverty reports has been the reduction of non-take-up of the benefit, continued energy supply, achieved by simplifying the process. Before automation, more than half of eligible households were estimated to miss out on support due to procedural complexity and the burden of providing documentation (Interoperable Europe YouTube channel, 2023[10]). The new system reduces the number of steps required from beneficiaries and frees up civil servants’ capacities to process more claims, making it easier for vulnerable households to access support.
The novelty of the system lies in its ability to integrate fragmented data sources through interoperability platforms and automate a multi-step eligibility process at scale. This represents a significant step toward proactive benefit delivery and provides the basis for further technical evolution with other AI models.
Standardisation across the region has been another key outcome. Providing the system through SaaS ensures that all municipalities use the same eligibility criteria and data sources, reducing disparities in implementation. This is particularly important in a context where, as AOC notes, municipalities previously applied the law in different ways due to unclear regulation and lack of guidance.
Challenges and limitations
The main obstacle to full automation, and consequently full coverage of eligible households, is legal. Current legislation requires explicit consent from beneficiaries, and sometimes from all household members, before municipalities can access the necessary data. To reduce friction in the process, AOC has proposed a shift to a presumed consent model with the possibility of opting out, applicable only in situations involving essential social benefits, with the goal of expanding coverage while maintaining the right to oppose to data processing or benefits provision, in accordance with data protection and social services laws.
Technical challenges related to data quality, standardisation, and sharing also remain an issue. Energy providers submit insolvency data to municipalities via spreadsheets, not following a specific format. For this reason, the system has been designed to process all the various formats available. AOC has proposed a secure data exchange interface, but its adoption by all the energy companies would be challenging.
The implementation of the AI solution also comes with risks in terms of transparency, especially for vulnerable households. Although a deterministic AI model is more explainable than a probabilistic one, the complexity of data flows and legal requirements can still make outcomes difficult to understand. AOC is developing algorithmic transparency reports and exploring digital consent mechanisms to address these issues (AOC, n.d.[11]), but public awareness of the automated process remains limited.
Transferability across administrations
The Catalonia case offers several takeaways for countries with decentralised and fragmented social security benefits provision, showing how central co-ordination is a key enabler in AI implementation. AOC’s role as a shared provider of digital infrastructure in Catalonia enabled the development of a solution that individual municipalities would not have been able to implement independently. This model is particularly relevant in systems where social security providers vary in size, capacity, and digital maturity, or where multiple actors intervene at various steps of benefits and services provision.
The case also demonstrates the importance of standardisation and data exchange infrastructure to support AI implementation. The use of the data exchange platform (Via Oberta), the social information system (Hestia), and a BPM solution made it possible to build a comprehensive and interoperable system using national data-sharing protocols. While many countries have some form of data interoperability infrastructure, Catalonia’s experience highlights the added value of real-time, secure access to data across institutions.
Ultimately, legal frameworks play a role in shaping the effectiveness of AI systems in the public sector. The current requirement for explicit consent in Catalonia limits the system’s potential for full automation and broader coverage. AOC’s proposed shift to a presumed-consent model with an easy opt-out option, applicable only in situations involving essential social benefits, could serve as a reference for other countries seeking to balance efficiency with privacy and ethical standards. In addition, AOC has produced a thorough algorithm transparency report1 aimed at ensuring explainability and understandability, improving accountability and oversight, and building trust while minimizing risks.
AI adoption has become widespread across Public Employment Services (PES), with many of them integrating it into core functions such as job matching and profiling (Brioscú et al., 2024[12]) (see Box 1.2). In Germany, the Federal Employment Agency (Bundesagentur für Arbeit, BA) has implemented an AI system to support the classification of job advertisements submitted by employers. This use case illustrates how machine learning can be applied to improve the efficiency of labour market services by automating the processing of unstructured data. The system reduces the manual workload of BA staff while maintaining human oversight and demonstrates how AI can be integrated into existing workflows to improve service provision without increasing the burden on external stakeholders.
AI is being used by half of Public Employment Services (PES) across OECD countries to enhance service delivery for jobseekers and employers. The most widespread application is job matching, where natural language processing and machine learning are used to compare jobseeker profiles with vacancies based on competencies. This allows PES to account for transferable skills and produce more accurate and personalised recommendations.
Algorithmic systems are also being used to support personalised and proactive guidance. Profiling tools use machine learning to identify people at risk of long-term unemployment early in their registration process, by analysing variables such as employment history and education. This enables PES to prioritise resources and provide tailored services such as training or counselling.
However, using AI in PES brings challenges that are critical given their impact on people’s access to services. Bias in profiling and matching tools can reinforce existing inequalities if algorithms replicate patterns from historical labour market data or rely on proxy variables like location or education. This can lead to unfair recommendations or misclassification of jobseekers, particularly those from vulnerable groups. To mitigate these risks, PES need systematic bias testing, transparent models, and human oversight to ensure that automated decisions remain fair.
Source: (Brioscú et al., 2024[13]).
Problem definition and institutional gaps
The BA receives over 100 000 job advertisements each year, sent by employers in different formats such as PDF attachments, or links to external websites via email. These submissions vary widely in structure and content, making it difficult to process them consistently. To be usable within the BA’s internal placement system, VerBIS, each advert must be standardised according to a set of predefined categories, including job title, contract type, required qualifications, and pay scale.
Previously, this classification was carried out manually by BA staff, requiring more than 10 minutes per advert on average. While increasingly requiring employers to submit job advertisements in a structured way partially solves the issue, the BA leaves the option of the email channel open in order to reduce the risk of reducing the number of offers shared.
AI system description
To automate the process of classifying job adverts, the BA developed ADEST (Automated Data Extraction and Structuring Tool), a machine learning system to extract and categorise information from unstructured sources. The tool proposes standardised job advert entries for the VerBIS system, the BA’s internal platform for job placement and counselling.
The BA ensured that employees retain decision-making authority. Case workers are shown on VerBIS the original job advert as it was received, alongside the various attributes prefilled by ADEST, including a confidence score for each calculated attribute (see Figure 1.2). Staff can accept, modify, or reject the AI’s suggestions, which also contributes to improve the model.
Technically, ADEST consists of several machine learning models – most notably optical character recognition (OCR) – that are trained on approximately 600 000 historical job advertisements. It was developed by the BA’s AI Competence Centre, with support from external contractors, and it is hosted on the internal infrastructure managed by the BA’s IT division. The user interface was designed to be intuitive and transparent, with confidence scores and traceability features that support explainability.
The ML model categorises the job posting, and the categories are shown to the user
Source: Bundesagentur für Arbeit.
Implementation conditions
The deployment of ADEST was enabled by a combination of organisational capacity, technical infrastructure, and internal governance mechanisms within the Federal Employment Agency. Development of the solution was led by the BA AI Competence Centre, which co-ordinated closely with business units across Germany throughout the design and testing phases. This collaboration ensured that the tool addressed a concrete operational need and could be integrated into existing workflows.
Legal compliance and ethical safeguards have been taken into account during the development process. The BA applies a standardised internal risk assessment to all AI projects, evaluating systems based on their level of automation and potential impact on users (see Box 1.3). ADEST was classified as low-risk due to its human-in-the-loop design and the absence of fully automated decision-making. The system was reviewed by the BA data protection office, the AI governance unit, and the IT security department to ensure compliance with the GDPR, the AI Act, and internal standards.
The implementation of ADEST also benefited from a favourable institutional context. The Federal government has expressed political support for the use of AI in public administration, particularly as a means to address workforce shortages in labour and social security services (Bundesregierung, 2024[14]). This strategic alignment has helped legitimise AI initiatives and provided a clear mandate for innovation.
To assess the risk factors associated with AI deployment, the BA developed the BA Data Ethics Guideline, which includes the following principles:
Human in the loop
We ensure that our algorithmic decision-making systems provide support in the interest of our customers (natural and legal persons) and our employees.
We ensure the autonomy of human decisions.
We take advantage of the opportunities of digitalisation in the interests of our customers and our employees and integrate their needs.
We encourage developments that support the meaningful work of our employees.
Robustness, safety and reliability
We ensure that the development and operation of our algorithmic decision systems meet current security standards and that users can rely on high-quality results.
We protect all information entrusted to BA.
We create robust algorithmic decision systems to minimise unintended consequences.
We use customer feedback to continuously improve our algorithmic decision systems.
Privacy and data quality
We ensure that personal data is processed in accordance with the data subject's wishes and in accordance with legal data protection requirements.
We protect personal data and maintain information self-determination.
We ensure compliance with data quality standards, thereby contributing to the best possible results of our algorithmic decision systems.
Transparency and explainability
We ensure that the interaction between humans and algorithmic decision systems is transparent and that the decisions made are user-friendly, understandable and explainable.
We educate our customers and employees about the goal we are pursuing with our algorithmic decision systems.
We ensure transparency within the BA on the development and use of algorithmic decision systems.
We show the opportunities and potential risks of algorithmic decision systems.
We inform clients and employees when interacting with algorithmic decision systems and results based on probabilities.
Equity, inclusion and diversity
We ensure that our algorithmic decision systems take into account human diversity and pay attention to fairness and gender equality.
We ensure that the principle of equal treatment is respected and that there is no discriminatory unequal treatment.
We ensure that our algorithmic decision systems promote gender equity.
We take into account all relevant stakeholders and involve them in the design process.
Using algorithmic decision systems, we actively promote accessibility and inclusion.
We systematically build feedback loops, actively capture user input and take it into account when evolving and developing our algorithmic decision systems.
We promote the data ethics competence and sensitivity of algorithmic decision system with developers and employees.
Social benefit and public interest
We balance economic, environmental and social interests to ensure that our algorithmic decision systems are used for the benefit of the general public and create value for society. The legally guaranteed protection of individuals and groups is of higher priority compared to the public good.
We promote social solidarity and improve opportunities for participation in the labour market and in vocational education and training.
We manage limited resources sustainable.
We promote developments that meet the needs of today's and tomorrow's generations.
We promote independent and evidence-based labour market research.
Accountability and legality
We are committed to developing and testing algorithmic decision systems in accordance with our ethical values and legal requirements, and we work closely with independent reviewers, academia and other organisations to do so.
We ensure accountability throughout the lifecycle of our algorithmic decision systems.
We document relevant decisions regarding the design and configuration of our algorithmic decision systems.
Source: Bundesagentur für Arbeit.
Outcomes and key lessons
The introduction of ADEST has led to measurable improvements in the efficiency and consistency of job advert processing within the BA. According to internal estimates, the time required to classify a job advert has been reduced by over 60%, compared to the previous manual process, which has freed up staff capacity for more complex activities.
The system has also contributed to greater standardisation in how job advertisements are recorded in the VerBIS system. Previously, classification practices varied across regional offices, depending on individual interpretation and workload. ADEST provides a consistent baseline by applying the same classification logic to all incoming adverts, which improves the quality and comparability of data used for job matching and labour market analysis.
While the system has been well received, the BA acknowledges that internal adoption has varied across teams. Training and change management efforts have helped increase take-up, but further work is needed to embed the tool into daily routines and ensure consistent use. The BA is monitoring usage patterns and gathering feedback to improve the system and to identify areas where additional training may be needed.
Challenges and limitations
The deployment of ADEST has shown several gaps related to transparency, explainability, and long-term governance. While users can review AI-generated classifications, the underlying model remains opaque. The BA has published documentation on the data sources, the performance metrics, and the model results on their intranet, but these materials are not accessible to external stakeholders. Moreover, there is currently no public registry of algorithmic systems in use, and the source code and training data are not disclosed, limiting external scrutiny and accountability.
The BA has also developed internal principles for trustworthy AI, covering transparency, fairness, and human oversight. However, these are not binding and apply only within the agency. Without alignment to broader OECD and other public sector standards, there is a risk of fragmentation in how AI governance is implemented across institutions, particularly when regulation evolves under the AI Act.
Finally, while ADEST was designed to minimise disruption by adapting to existing workflows, this approach also limits the scope for more transformative change. While the system was not designed to address upstream issues such as the heterogeneity of job advertisements or the lack of structured data from employers, it is still able to improve the efficiency of specific tasks. As with many AI applications in the public sector, the risk is that automation reinforces existing inefficient processes rather than redesigning them entirely.
Transferability across administrations
The ADEST system demonstrates how AI can be deployed effectively in public administration by targeting a specific, time-consuming task without requiring structural changes. This approach is particularly relevant for institutions that, like the BA, receive large volumes of unstructured data and cannot easily impose standardisation on external actors.
The German case also underscores the importance of internal adoption and change management. AI systems require sustained investment in training and communication to ensure consistent use. This is a shared challenge across other social security institutions that are still working to build AI literacy among staff and embed AI tools into daily operations, as showcased in Chapter 3. The BA’s experience suggests that replicability depends not only on technical feasibility, but also on organisational readiness and sustained institutional support.
Kela, Finland’s national social security institution, has developed an AI platform to improve internal operations and reduce administrative burden. The platform automates the processing of user-submitted documents and messages, leading to faster and more consistent service delivery. While the system does not determine eligibility or make benefit decisions, it improves the efficiency of claims processing and supports staff in managing high volumes of data. This use case illustrates how AI can be applied to procedural tasks in a low-risk way, aligned with goals of trustworthy AI in public administration.
Problem definition and institutional gaps
Kela faces operational inefficiencies due to the manual handling of millions of documents and customer interactions each year. Most benefit applications are submitted via the online platform OmaKela, and in many cases they require additional documentation not necessarily retained by public institutions, such as rental contracts or bank statements, that must be manually reviewed and classified. Similarly, support messages and phone calls with customers often contain unstructured information that needs to be categorised or transcribed, adding to the workload and delaying processing times.
Beyond these operational challenges, Kela must balance efficiency with transparency and oversight. While the institution benefits from a high level of digital maturity, many processes require human involvement. This mirrors the situation faced by the German Federal Employment Agency, where the lack of tools to process job offers limited responsiveness. In Kela’s case, the direct link between document processing and benefit delivery increases the pressure and the oversight on the procedures.
Manual procedures also affect the user experience. Complex steps, delays, and inconsistent communication can discourage vulnerable individuals from completing their applications, contributing to non-take-up. Kela’s AI platform was developed to address this issue by automating routine tasks while preserving human oversight. Rather than replacing decision-making, the system supports it, streamlining workflows without compromising accountability.
AI system description
The AI platform automates the intake and handling of documents attached to benefit applications. It performs a series of tasks such as format conversion, text recognition, image correction, barcode reading, and classification. These steps are executed in a single workflow and applied across multiple benefit types.
Beyond the processing of attachments, the platform includes tools for transcribing customer phone calls and classifying support requests to transfer them to the appropriate business unit. Phone call transcription is not fully automated: staff can choose whether to activate it and select among different speech-to-text models depending on their quality.
The platform runs on an on-premises cloud infrastructure fully managed by Kela, using modular components. It relies on open-source software such as Tesseract for OCR, Whisper for speech recognition, and PyTorch for classification. These tools are containerised and orchestrated through Kubernetes, which means that applications can run consistently across different environments with automated management. This allows for scalability and resilience.
Implementation conditions
Kela’s AI implementation strategy has been based on long-standing institutional choices that prioritise autonomy and sovereignty. The decision to develop and operate the AI platform in-house, using open-source technologies on a private cloud infrastructure, ensures full control over the system, reduces dependency on third-party vendors, and allows for incremental adaptation over time.
The use of open-source software also supports recruitment and knowledge sharing. By aligning with widely used technologies and contributing to open-source communities, Kela can attract talent with relevant expertise and benefit from a broader ecosystem of support.
Beyond technical implementation, Kela has worked closely with other public authorities to shape its AI governance strategy. National supervisory authorities have helped by issuing guidance on responsible AI use. This collaborative approach has been reinforced by strong political support for AI in public administration, creating a favourable environment for experimentation within clear regulatory boundaries.
Outcomes and key lessons
The AI platform has delivered measurable improvements in Kela’s internal operations. In 2024, it processed over 16 million attachments, saving an estimated 38 person-years of staff time. By automating routine tasks, the system has reduced manual workload, improved processing speed, and freed up capacity for more complex interactions with beneficiaries.
Efficiency gains have also translated into infrastructure savings. The platform’s ability to resize and convert documents reduced storage needs by 24 terabytes, lowering operational costs and reducing the environmental footprint of the infrastructure.
While the system does not directly address benefit eligibility or non-take-up, it strengthens the foundation for future improvements. By streamlining back-office processes, Kela is better positioned to explore more proactive service models, including automated outreach and real-time eligibility checks—provided that legal and ethical safeguards remain in place.
Challenges and limitations
The effectiveness of language-processing tools such as speech recognition and message classification is limited by the availability of high-quality training data in Finnish, and the near absence of data in Sámi languages spoken in northern Finland. This affects accuracy and requires staff to remain involved in reviewing outputs.
Maintaining full control over the system also comes with trade-offs. Kela’s commitment to in-house development and on-premises infrastructure requires sustained investment in skills and resources. As the organisation transitions to a hybrid cloud model, it will need to balance flexibility with its emphasis on sovereignty and data security.
Transferability across administrations
Kela’s case shows how AI can be introduced in social security institutions without fully automating decisions. In a field where administrative actions affect people’s lives, legal and institutional frameworks place strict limits. By focusing on document classification and message handling, Kela demonstrates how AI can improve efficiency while respecting the safeguards that ensure fairness and accountability.
The choice to build and maintain the platform in-house, using open-source technologies, is especially relevant for institutions with strong internal technical capacity. This approach reduces dependency on external vendors, lowers long-term costs, and allows the institution to adapt the system as needs evolve.
At the same time, the case highlights the importance of institutional readiness. Kela’s success depended on more than infrastructure: it required a clear strategy, stable governance, and a culture that values collaboration between technical and operational teams. These conditions are not always present in more fragmented systems, where responsibilities are decentralised and digital transformation is driven by procurement cycles rather than long-term planning.
The three use cases presented in this chapter show how AI can support social security institutions in addressing operational inefficiencies, improving service delivery, and laying the groundwork for more proactive benefit provision. However, successful implementation also depends on the institutional context, the legal framework, and operational priorities.
Common enablers: clear use cases, shared infrastructure, and institutional support
All three cases emerged from a clearly defined problem, where the benefits of adopting AI outweighed the risks. In Catalonia, the heterogeneity of capabilities across municipalities and the repetitive task of data verification highlighted the need for a centralised solution, enabled by the energy poverty law. In Germany and Finland, the demand for automation came from the business units, while AI governance structure at the organisation level provided clear guardrails. In each case, AI was not chosen by default but selected as the most appropriate tool for the task.
In-house skills and technical infrastructure were shared enablers. Each institution built on existing digital systems: shared interoperability platforms in Catalonia, internal IT infrastructure in Germany, and the on-premises cloud in Finland. The AOC, the BA, and Kela all benefited from internal skills and expertise. These foundations allowed AI systems to be integrated into existing systems and workflows, rather than requiring the development of a custom-made solution, or reliance on external contractors.
Ultimately, organisational and institutional support was critical. Whether through the central role of AOC in Catalonia, co-ordination across internal teams in the BA, or in-house technical capacity within Kela, each institution ensured that the AI development aligned with broader strategies and policy objectives.
Recurring challenges: transparency, bias, and skills
A lack of algorithmic transparency was the main limitation across all three cases. While each system has been designed with safeguards, such as deterministic logic in Catalonia or human-in-the-loop validation in Germany and Finland, none of the institutions maintained a public registry of AI systems or published code or detailed documentation. This limits accountability and external scrutiny, making it hard to fully understand or challenge outcomes.
Bias and fairness were also concerns, particularly in systems that rely on historical data or unstructured inputs. While none of the use cases involved fully automated decisions, all three institutions recognised the risk of reinforcing existing inequalities if AI systems are not carefully monitored. In the absence of clear documentation or explainability mechanisms, these risks are harder to detect and mitigate.
Another challenge for AI adoption is the lack of high-quality training data in all relevant languages. Language-processing tools such as speech recognition or message classification require large datasets, which are scarce for Finnish and almost absent for Sámi languages, reducing accuracy and requiring human review. Similar issues arise in Catalonia, where the coexistence of Catalan and Spanish complicates the development of robust language models.
Skills gaps were a shared constraint, though addressed in different ways. Germany and Finland invested in internal capacity, while Catalonia relied on a central digital solutions provider to support municipal service delivery. In all cases, staff training and organisational change were essential to ensure adoption and oversight. This topic will be further explored in Chapter 3.
AI has the potential to improve efficiency, reduce burdens on both staff and citizens, and enhance access to benefits and services in the area of social security as well as other public services. AI tools can help to reduce the administrative burden for both applicants and public sector employees and improve access to services, for example, by performing automated eligibility checks or facilitating document processing. This is reflected in three use cases from the EU:
Catalonia: AI for checking eligibility of means-tested benefits
In the Spanish region of Catalonia, an AI system was developed to facilitate the implementation of a law to tackle energy poverty, by automating eligibility assessments. AI is used within a cloud-based platform to generate energy poverty reports based on integrated data from multiple public sources. This solution significantly reduces the administrative burden on municipalities and citizens. However, full automation is constrained by legal requirements for explicit consent, and data-sharing processes still rely on manual inputs from private energy companies. Catalonia’s experience highlights the critical importance of data interoperability, shared infrastructure, and the role of co-ordination between stakeholders in enabling scalable AI adoption.
Germany: machine learning for document recognition and classification
Germany’s federal employment agency (BA) implemented a machine learning tool, ADEST, to classify unstructured job posting data for integration into its central employment system. The AI system significantly reduces manual effort by processing over 100 000 annual job postings, presenting structured data to civil servants along with confidence scores, for validation. The BA’s approach illustrates how AI can improve service responsiveness while maintaining trust and oversight.
Finland: AI to simplify claims processing
Finland’s social security institution, Kela, operates an in-house AI platform to automate the processing of documents and support requests. This platform automates the processing of over 16 million documents annually, supporting applications submitted through its online portal. The system performs tasks such as format conversion, classification, OCR, and document resizing—saving an estimated 38 person-years of work and over 24 TB of storage annually. Finland’s approach shows the long-term value of maintaining in-house infrastructure and FOSS tools to retain technical sovereignty, manage costs, and build institutional expertise.
From this, there are three key takeaways for the social security sector:
Enablers matter: These cases underline the importance of foundational investments in digital infrastructure, data governance, and in-house institutional capability. Centralised digital platforms can accelerate uptake and lower the entry barrier for decentralised administrations through reduced costs.
Guardrails are necessary: Strategic, effective, responsible, and trustworthy AI deployment requires human oversight, data protection, transparency, and fairness. Germany’s model emphasises ethical AI by design, while Finland’s infrastructure reflects strategic priorities.
User-centred engagement is critical: All three examples highlight that AI systems must be tailored to administrative workflows and user needs. This is because effective solutions need end-users – both internal and external – to be actively involved in design and iteration.
Overall, these examples demonstrate that AI, when carefully implemented, can reduce barriers to access, increase efficiency, and build more inclusive and responsive social protection systems.
References
[11] AOC (n.d.), Algorithmic transparency: Automation of the energy poverty report, https://www.aoc.cat/en/ia-transparencia-pobresa-energetica/ (accessed on 21 July 2025).
[8] AOC Consortium (2023), Pilot for the automation of the energy poverty report to avoid supply cuts, https://www.aoc.cat/en/blog/2023/pilot-pobresa-energetica/ (accessed on 23 June 2025).
[12] 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.
[13] 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.
[14] Bundesregierung (2024), Antwort 20/12831 - auf die Kleine Anfrage - Drucksache 20/12516 - Jobcenter und Arbeitsmarktförderung nach dem Zweiten Buch Sozialgesetzbuch, https://dserver.bundestag.de/btd/20/128/2012831.pdf (accessed on 1 August 2025).
[9] Catalunya (2017), LLEI 12/2007, d’11 d’octubre, de serveis socials, https://portaljuridic.gencat.cat/ca/document-del-pjur/?documentId=415692 (accessed on 21 July 2025).
[6] Catalunya (2015), LLEI 24/2015, del 29 de juliol, de mesures urgents per a afrontar l’emergència en l’àmbit de l’habitatge i la pobresa energètica.
[10] Interoperable Europe YouTube channel (2023), SEMIC 2023: Parallel Track on Trustworthy AI for Interoperability in the Public Sector, https://www.youtube.com/watch?v=Tg_vVoDo6wc&t=1016s (accessed on 21 July 2025).
[4] OECD (2025), “AI and the future of social protection in OECD countries”, OECD Artificial Intelligence Papers, No. 42, OECD Publishing, Paris, https://doi.org/10.1787/7b245f7e-en.
[3] OECD (2025), Governing with Artificial Intelligence: The State of Play and Way Forward in Core Government Functions, OECD Publishing, Paris, https://doi.org/10.1787/795de142-en.
[2] OECD (2024), Megatrends and the Future of Social Protection, OECD Publishing, Paris, https://doi.org/10.1787/6c9202e8-en.
[1] OECD (2024), Modernising Access to Social Protection: Strategies, Technologies and Data Advances in OECD Countries, OECD Publishing, Paris, https://doi.org/10.1787/af31746d-en.
[5] OECD (2022), “OECD Framework for the Classification of AI systems”, OECD Digital Economy Papers, No. 323, OECD Publishing, Paris, https://doi.org/10.1787/cb6d9eca-en.
[7] Pye, S. et al. (2017), “Energy poverty across the EU: Analysis of policies and measures. Analysis of policies and measures”, in Europe’s Energy Transition: Insights for Policy Making, https://doi.org/10.1016/B978-0-12-809806-6.00030-4.