Governing with Artificial Intelligence

AI in public service design and delivery

The development and use of AI has permeated public service design and delivery in recent years, with 67% of OECD countries using AI to improve this function (OECD, 2024[188]). Public servants are increasingly taking it upon themselves to experiment with AI in their job duties (Bright et al., 2025[189]). In general, the aim is to make them more efficient, effective and responsive to the public’s needs. AI can help achieve this thorough the automation of processes, more efficient allocation of resources (e.g. optimising staff levels), supporting decision-making, improving citizen participation, and tailoring and personalising content and service pathways for citizens (2024[60]; 2023[190]; 2019[191]; OECD, 2025[192]). AI can also increase governments’ understanding of socio-economic challenges service users are facing —by picking up on trends and causal relations that are not detectable by humans — thus, improving the responsiveness and personalisation of public services (Anshari et al., 2024[193]). This can allow governments to design novel public services that are more proactive in anticipating and responding to user needs (Madan and Ashok, 2023[194]).

In certain circumstances, automated decision-making in public services may also be perceived as fairer and less prone to human error or corruption (Miller and Keiser, 2020[195]; Leyer and Schneider, 2021[196]). However, a variety of human behaviour factors — how people perceive, trust and interact with AI-driven decisions — will determine its effectiveness and legitimacy. Better service delivery and service effectiveness can boost citizens’ satisfaction and their trust in government (Sun and Medaglia, 2019[197]; OECD, 2024[46]). Some argue that this moment requires a broad rethinking of how impacts AI impacts public administration structures underlying public service delivery, beyond focusing on efficiency gains (Babšek et al., 2025[198]).

Current state of play

Automating and optimising tasks

The most common tasks in public services are routine — recording, preparing, sorting, classifying, filing and verifying the accuracy of information — and have a high potential to be automated (Straub et al., 2024[199]). AI can streamline bureaucratic tasks, freeing time for public servants to focus on those tasks requiring human judgement, creativity, discretion and decision-making. Real-world use cases highlight the importance of cost efficiency in integrating AI applications to service delivery. For example, Brazil employs a RPA solution to detect and dismiss predatory and repetitive legal claims —claims engineered by wealthier entities to burden smaller defendants with exorbitant, extraneous data requests and costs — thereby streamlining court procedures and conserving valuable time.1 In Spain, RPA technology automates or accelerates low-value repetitive tasks in public services, ensuring an efficient distribution of non-contributory disability and retirement pensions.2 Additional examples in Austria and Spain can be seen in Box 5.42.

Box 5.42. AI solutions in planning processes

Vienna’s digital building permit reviews (Austria)

The Building Regulations Information for Submission Envolvement (BRISE-Vienna) in Austria introduces a digital innovation in building permit reviews, transforming traditional, paper-based processes into a comprehensive digital workflow. In the BRISE project, AI-based verification routines automatically compare the digital 3D Building Information Models submitted by planners with a municipal digital reference model that encapsulates regulatory requirements. This automated process not only quickly identifies any deviations from established standards but also provides immediate feedback, significantly reducing the traditionally lengthy review period.

Automating the energy poverty report in Catalunya (Spain)

The Automating Energy Poverty Report in Catalunya, in Spain, is an initiative by the Open Administration Consortium of Catalonia (AOC) to develop an AI-driven platform for automating the generation of energy poverty reports. This platform targets the manual and bureaucratically expensive process of determining social vulnerability and producing energy poverty reports for municipalities. The innovation lies in automating the energy poverty report generation process, which traditionally involved manual, time-consuming and error-prone methods. The initiative addresses the problem of social vulnerability determination by using robotic process automation (RPA), business process modelling (BPM), and AI technologies to streamline the process. Previously, municipalities handled this task independently, leading to inconsistent results and uneven resource allocation. The automation of energy poverty reports leads to several benefits, including centralised AI-powered support, seamless integration with diverse information systems, and enhanced interoperability among companies, municipalities and other administrations. Overall, the platform enables more efficient and accurate social vulnerability determinations, streamlined administrative processes, and equitable resource allocation for energy poverty mitigation.

Automating property records in Greece

The Hellenic Cadastre, previously hindered by manual paper processing that delayed property transactions for months or even years, has introduced an AI-driven system to automate the reading, categorization, and legal assessment of property contracts. This innovation has cut assessment times from several hours to under 10 minutes and reduced costs from EUR 15 to just EUR 0.11 per case. It has also enhanced legal security for property owners and boosted economic activity by accelerating transaction completion.

Source: https://oecd-opsi.org/innovations/brise-vienna-building-regulations-information-for-submission-envolvement, https://uia-initiative.eu/en/news/piloting-brisevienna-results-journal-ndeg-3, https://www.aoc.cat/blog/2023/pilot-pobresa-energetica., https://www.ekathimerini.com/economy/1245436/smart-solutions-for-land-register, https://www.microsoft.com/en/customers/story/20308-hellenic-cadastre-azure-open-ai-service, Government of Greece officials.

Automation is relatively easy with AI when dealing with repetitive and low discretionary tasks if the data quality is assured. Automation becomes more difficult when higher discretionary tasks are involved but used to uncover new options, detect anomalies and risks and better plan services. However, potential challenges for ensuring systems are trustworthy need to be considered, as seen in Box 5.43.

Box 5.43. Automation for smoother delivery of services in Sweden

The Public Employment Service in Sweden rolled out Prepare and Match across the country in 2021 to improve consistency and accuracy of labour-market related assessments, and thereby improve the efficiency of resource allocation. Enrolled jobseekers get support in the form of training or guidance from a chosen provider. Decisions about whether jobseekers should be approved to participate in Prepare and Match are assisted by a decision-support tool called BÄR.

BÄR is a ML system trained on historical data consisting of 1.1 million profiles collected over a period of 10 years. It gives different recommendations for action based on the jobseeker's probability of finding a job within six months. Factors considered by the system include age, sex, education, residential location (and associated demographic information), and previous (un)employment activities.

Decisions concerning Prepare and Match are formally made by caseworkers, who are instructed in guidelines to primarily adhere to the automated recommendation. Overriding a negative recommendation from the system is difficult and requires contacting a special working group within the agency. The role given to the AI in the decision-making process highlights the potential impact of AI for jobseekers while underscoring challenges with partially or fully removing them from decisions that affect them. Jobseeker's own judgment concerning their need for support is not considered by the decision-support system

Source: (Berman, de Fine Licht and Carlsson, 2024[200]; OECD, 2024[201]).

Offering new sources of information for advanced insights

AI can help to maximise the quality and utility of data, as well as humans’ and machines’ ability to process and analyse it (Jarrahi et al., 2023[202]). This can contribute to AI being used for more advanced and real-time impact analysis of public services. Able to recognise patterns that may be undetectable by humans, AI applications can quickly identify problems that may have gone unnoticed (Gesk and Leyer, 2022[203]). They can also process data from a variety of structured and unstructured sources (e.g. images, text, meta-data), making information management much easier (Cockx, Lechner and Bollens, 2023[204]). In Finland, the city of Tampere collects data from traffic cameras to feed a predictive algorithm to predict visitor flows in the city centre, helping businesses and public servants make informed decisions, representing an entirely new public service (Box 5.44).

Box 5.44. AI-powered visitor flow forecasting with Tampere Pulse in Finland

Tampere Pulse is an AI-driven predictive analytics service that forecasts visitor flows in the city centre of Tampere, Finland. Developed in response to challenges faced by local businesses in predicting customer traffic, the service uses data from the city's Internet of Things (IoT) platform, including machine vision-equipped traffic cameras, weather conditions and event schedules. The result is a month-long visitor flow forecast, accessible on both mobile and desktop devices through an intuitive map-based interface.

The service has demonstrated significant impact, with a prediction accuracy of 79.33% for the exact visitor flow category and 97.55% for the correct or adjacent category up to three weeks in advance. This precision allows businesses to optimise workforce planning, inventory management and operating hours, leading to cost savings and improved customer service. User feedback has been highly positive, with an ease-of-use rating of 4.47 out of 5 and a recommendation likelihood of 7.16 out of 10.

Beyond business benefits, Tampere Pulse helps city residents and visitors plan their movements and assists maintenance and security teams in resource allocation. Future enhancements aim to expand coverage, integrate purchasing behaviour insights and link with enterprise resource planning (ERP) systems. The initiative showcases the potential of AI and IoT in fostering a data-driven, smart urban environment while serving as a model for replication in other cities worldwide.

Source: https://oecd-opsi.org/innovations/tampere-pulse.

These new advanced insights often result from a combination of information from variety of sources and technologies. In Portugal, the Land Cover Monitoring System uses AI to integrate stakeholder needs with satellite technology to create an open cartography, promoting efficient, accessible and transparent insights into land cover (Box 5.45). The Saudi Data and Artificial Intelligence Authority (SDAIA) developed Smart C to support data-driven urban planning and real-time decision-making, enabling tools such as traffic simulation, environmental monitoring, and AI-based visual pollution detection (SDAIA, 2024[205]).

Box 5.45. Portugal’s Land Cover Monitoring System (SMOS)

The Portuguese Land Cover Monitoring System (SMOS) is a public service that provides open-access, AI-powered land use and land cover maps to support evidence-based decision-making. Developed by the Directorate-General for Territory (DGT), SMOS ensures that policymakers, researchers and the public have reliable data on land changes, enhancing transparency and sustainable land management.

SMOS analyses Sentinel-2 satellite images, high-resolution orthophotos and official geographic data using AI to produce accurate cartographic products, such as Land Use and Land Cover Maps (COS), dynamic Conjunctural Land Cover Maps (COSc), and specialised monitoring products such as the Annual Map of Temporary Agricultural Crops (MACAT). These products are based on automated classification of satellite imagery into thematic land cover categories, using AI to distinguish features such as urban areas, forests, water bodies and agricultural fields.

SMOS has progressed through three phases: initial research and development; engagement with users to align products with public needs; and now full operational deployment. Since 2022, SMOS has delivered annual Land Use and Land Cover Maps (COS), dynamic Conjunctural Land Cover Maps (COSc), and specialised monitoring products, all accessible via Portugal’s National Spatial Data Infrastructure (SNIG).

Source: https://smos.dgterritorio.gov.pt.

Providing constantly accessible information and services

Virtual assistants and chatbots are among the most common uses of AI in public services (Ramires Hernández, Valle-Cruz and Mendoza Méndez, 2022[206]; Chen, Gascó-Hernandez and Esteve, 2023[207]). These tools give public service users access to information quickly, 24 hours a day. Currently, chatbots range from basic (information retrieval-based algorithms) to more advanced chatbots that use generative AI systems, such as LLMs, and tap into real-time service-related data. The latter can respond to queries with personalised messages, diverging from scripted conversations to tailor information and recommendations to users’ specific needs, including plain language or sign language based on text or speech input. In addition to being convenient for users, chatbots can make public services more accessible, increase the uptake of services and boost engagement.

Efforts in Portugal serve as an example of how governments can integrate the latest foundation models into public services (Box 5.46). In the United States, the Department of Homeland Security’s Emma chatbot provides users with information on immigration services, green cards and passports (Chen, Gascó-Hernandez and Esteve, 2023[207]). In Singapore, the chatbot Gov.sg allows users to file and track the status of complaints about public services. In Brazil, the municipality of Recife allows for over 700 public services in the “Conecta Recife” platform to be automatically distributed and delivered to its citizens, with a chatbot leveraging AI systems to immediately provide information to users. In Chile, a chatbot was piloted on ChileAtiende to respond to citizen queries about public services, demonstrating the potential to provide personalised assistance and reduce response times across millions of requests.3

Beyond direct interactions with citizens, AI-powered chatbots are also being used within public administrations to support staff in delivering services more efficiently. For example, Caddy, developed by the United Kingdom’s i.AI, acts as a copilot for customer service agents, helping them quickly retrieve reliable information from government resources while ensuring accuracy through a human-in-the-loop validation system.4 Similarly, Singapore’s AIBots platform allows government teams to customize AI assistants for internal processes, streamlining workflows and improving decision-making.5 In France, a chatbot was developed to support public servants to ensure better delivery of services across different fields (Box 5.46). The government of the United Arab Emirates has sought to innovate the virtual assistant concept by representing both virtual and human assistants in the form of a hologram (Box 5.47). In focusing on a specific area, the Israel Planning Authority — the national body responsible for overseeing spatial planning, infrastructure and housing developments — in May 2025 commenced proof of concept for a chatbot designed to help planning professionals in reviewing plans against national requirements, streamlining lengthy review processes, which often take years, ensuring compliance with legal and professional standards.

Box 5.46. Chatbots that streamline access to information

Portugal’s chatbot provides information to citizens

In May 2023, Portugal’s Administrative Modernisation Agency (AMA) launched a chatbot with a realistic avatar to assist citizens with questions about the Mobile Digital Key (MDK). Developed in partnership with Microsoft, DareData Engineering and Defined.ai, the chatbot is based on Azure OpenAI’s GPT-3.5 turbo model and can recognise and reproduce both text and voice in Portuguese. From May to December 2023, it facilitated 23 780 conversations (averaging 101 per day) and contributed to a 10% increase in MDK activations.

In January 2025, AMA expanded the chatbot into a full-fledged virtual assistant on the gov.pt portal, reinforcing Portugal’s commitment to digital inclusion and public service transformation. Now powered by ChatGPT with a specialised knowledge base, it provides information on over 2 300 public services and offers support in 12 languages, eliminating language barriers for citizens of diverse nationalities. Integrated into over 5 000 pages of gov.pt, the virtual assistant is set to introduce new transactional features in 2025, including process tracking and appointment scheduling, further enhancing its role as a key public service tool.

France’s Albert supports public services employees

In April 2024, the French government introduced Albert, a sovereign generative AI developed by the DataLab of the Direction Interministérielle du Numérique (DINUM) to assist public administration agents in responding to citizens' requests more efficiently. Trained on open-source models and enriched with practical guides from service-public.fr, Albert helps government employees search regulations, draft summaries, and provide accurate information to citizens. Albert is based on open-source models from Meta (Llama) and Mistral. As a digital common, the Albert application programming interface (API) provides open, reusable generative AI models, lowering the adoption threshold for public agencies and contributing to a national digital public infrastructure for AI.

In early 2024, Albert had been tested by 60 agents across 30 France Services centres, offering suggested responses, verified sources and useful references — while keeping human agents in control of final interactions. Officially launched nationwide in April 2024, Albert is set to expand across public administration, aiding tax services, environmental project reviews, judicial transcriptions and more.

France’s Sofia improves public access to information

In 2025, the French Ministry of Ecological transition released Sofia, a conversational agent that facilitates access to the organisation’s scientific and technical knowledge. Sofia enables users to query a specialised, pooled documentary database, making information searches quicker and more efficient. The system is based on a LLM developed by Mistral and can provide relevant answers based on users’ needs. Sofia has been developed with attention towards the ecological impacts. For this reason, it was decided to show users the carbon footprint of each search.

Greece’s AI assistants help users navigate and find information

Greece’s mAigov is an AI-powered assistant on the gov.gr portal that helps users navigate over 1 300 digital services and over 3 200 administrative procedures by answering questions in natural Greek via text or voice. mAigov can understand and respond to queries submitted in natural language, providing an intuitive way for users to locate the services they need. The chatbot incorporates reasoning capabilities and maintains the context of conversations to provide optimal user experience. Accessible via both mobile phones and computers, mAigov was initially launched in a pilot mode in December 2023 and can handle up to 240 concurrent conversations in Greek. Its functionality has since been expanded to support 25 languages, catering to a broader range of users, including Greek citizens abroad. With over 1.6 million queries answered, mAigov has proven to be a valuable tool in enhancing citizen interaction with the state. Future developments for mAigov include the ability to issue certificates and schedule appointments on behalf of citizens, further streamlining public service delivery. In addition, the Hellenic Ministry of Tourism, in collaboration with the Ministry of Digital Governance, has developed mAiGreece, a digital AI assistant to enhance travel experience in Greece. The app provides administrative and health & safety information, and recommendations for exploring Greece.

Source: https://oecd-opsi.org/innovations/virtual-assistant-for-public-services; https://digital.gov.pt/noticias/assistente-virtual-do-gov-pt-facilita-a-interacao-em-12-idiomas, https://www.info.gouv.fr/actualite/ia-connaissez-vous-albert, https://www.lefigaro.fr/secteur/high-tech/albert-l-ia-generative-de-l-etat-desormais-deploye-a-grande-echelle-20250212, https://digital-skills-jobs.europa.eu/en/latest/news/maigov-new-ia-digital-assistant-govgr-service-citizens, Government of Greece officials.

Box 5.47. United Arab Emirates Smart Mission for consular services

Implemented in line with the UAE Strategy for Government Services and the National AI Strategy 2031, the Smart Mission is designed to streamline and digitalise consular services, significantly reducing processing times, improving service delivery and enhancing user experience. At the core of this transformation is the AI-powered Document Attestation Service, which uses an AI-driven attestation verification system to authenticate official documents by recognising official stamps of personal and commercial documents. This ensures higher accuracy in attestation processes and enhances security measures to protect against fraud and misuse.

Complementing this is the AI-Driven Holographic Assistance, an interactive AI virtual assistant that can engage with users directly through a hologram. These AI agents support users throughout the user journey, answering queries instantly in multiple languages and assisting with services. If further assistance is required, the AI assistant can transfer the conversation to a human consular officer in holographic form. Predictive AI systems analyse user behaviour, historical data and interaction patterns to provide proactive consular support. This allows the system to anticipate frequent requests, suggest relevant services and notify users of required actions ahead of time.

Finally, an AI-Powered Facial Recognition is also provided as an optional service for seamless access and proactive services, where users can authenticate their identity using facial recognition. This eliminates the need for physical documents and allows for tailored assistance without manual input, such as for the issuance of Return Documents for UAE nationals with lost or damaged passports, without any documentation needed.

Source: Government of UAE officials, https://u.ae/en/about-the-uae/digital-uae/digital-transformation/platforms-and-apps/smart-mission.

Recent research also explored the role of chatbots in behavioural interventions (e.g. during public health and risk communications during the COVID-19 crisis), such as increasing the willingness to socially distance. For example, a randomised controlled study found that an AI chatbot, designed with behavioural science, improved understanding of COVID-19 guidelines and increased intentions to seek testing services by reducing uncertainty around protective behaviours. By using animations and engaging messaging, the chatbot effectively addressed users’ informational and emotional needs, demonstrating how AI can shape decision-making and promote public health behaviours (van Baal, 2024[208]).

Chatbots are often designed primarily to streamline operations, rather than to optimise user experience for citizens. However, research shows that users engage more positively with AI-driven services when chatbots maintain continuity in identity and retain user information across interactions. Migratable AI systems, which preserve both an agent’s identity and user history across different interactions, have been found to increase trust, competence and social presence (Tejwani et al., 2020[209]). This suggests that designing chatbots with a stronger emphasis on familiarity and user continuity — not just efficiency — can significantly improve engagement. Applying behavioural insights can further enhance their effectiveness by ensuring chatbots detect when users need human support, rather than frustrating them with prolonged automated interactions. By recognising signals of confusion or urgency, AI chatbots can seamlessly escalate complex issues to human agents, reducing friction and improving service accessibility. Prioritising continuity and user-centric design in chatbots not only improves efficiency but also fosters trust, accessibility and engagement in public services.

Improving personalisation and tailoring of services

AI use for public service delivery also enables a shift from top-down implementation of public services to approaching design and delivery based on user needs. By enabling the customisation of services, AI enhances the relevance and responsiveness of interactions with each citizen, making services more accessible and significantly boosting public engagement (OECD/UNESCO, 2024[131]). For instance, France is using AI to enhance public service delivery, ensuring a smoother relationship between users, their individual needs and public servants (Box 5.48).

Box 5.48. France’s AI-assistance for personalised public service delivery

In France, the Direction Interministérielle de la Transformation Publique (Interministerial Directorate of Public Transformation - DITP) has launched an innovative AI-powered platform called Services Publics+ to improve the quality and efficiency of public services nationwide. The platform, which involves 45 million citizens and 2.5 million public servants, incorporates three main AI components: speech-to-text functionality for improved accessibility; automated assistance for civil servants in drafting responses; and real-time analysis of user feedback to identify trends and pain points.

The system’s implementation maintains a careful balance between automation and human oversight. Public servants retain full control over the AI-generated responses, with the ability to modify or reject suggestions before sending them to citizens. In line with EU AI Act principles, citizens are notified when responses have been assisted by AI, and all AI-generated answers are published as open data to ensure transparency. The system has been deployed across 12 ministries and 40 public services, with more than 4 000 civil servants using it regularly.

The impact of this innovation has been significant, with average response times dropping from 19 days to three days. The system has achieved high satisfaction rates, with 78% of public servants finding the AI suggestions useful, and 70% of AI-suggested elements being validated by public servants. Most importantly, citizen satisfaction with responses has increased, with 68% of users finding AI-assisted answers useful, compared to 57% for traditional responses. This demonstrates how AI can be effectively integrated into public services while maintaining human expertise and improving both efficiency and user satisfaction.

Source: https://www.plus.transformation.gouv.fr, presentation by France officials at AI Action Summit (February 2025).

Beyond their use in chatbots, generative AI systems (such as text, image, video, code or audio generators) can further personalise services for citizens. Capable of building user profiles and micro-sorting citizens into specific types of user groups, AI can tailor information and the services themselves to the complex needs of users and groups of users, including those in vulnerable and disadvantaged conditions. This can, for instance, be used to ensure eligible individuals are aware of social programmes and benefits available, event potentially pre-emptively providing them (OECD, 2024[210]) (Box 5.49). It can also tailor services around life-events by anticipating context-dependent service needs in a range of fields, including education, transportation and employment (Kopponen et al., 2024[211]). It can also assist in public employment services, such as by matching employers and job seekers (Box 5.50).

Box 5.49. How social protection systems are embracing AI

Advances in data and technology are helping to improve the accessibility, coverage and delivery of social benefits and services, though non-take-up remains a significant issue. Recent advances centre around linking administrative data from different sources to improve enrolment in social programmes. By linking datasets, governments can streamline application and renewal processes, inform potential beneficiaries of their eligibility for benefits, and (in limited cases) automatically enrol eligible users.

Despite its potential, advanced uses of AI have been little used by social affairs ministries, including for programme take-up. The most common uses of AI in social programmes in OECD countries have been automated client support (e.g. chatbots and digital assistants), automation of internal processes (e.g. data entry and document processing) and fraud detection and finding anomalies in benefit claims.

Despite this limited usage, AI offers significant promise in social protection, including through the identification and enrolment of service users, enhancement of user support, and improving the efficiency and timeliness of benefit and service delivery in response to changing needs. Emerging uses include:

• advancing current uses of chatbots, automation of routine tasks and fraud detection

• improving client service by accompanying human service providers in sensitive situations

• better automating benefit decisions based on statutory eligibility rules

• new types of predictive analytics, such as identifying recurring risks and vulnerabilities to target social programme interventions to at-risk individuals and groups.

Perhaps the most rewarding potential use case for AI in social protection is identifying high-risk beneficiaries and proactively reaching out to them — or automatically enrolling them — to improve social protection coverage. This is starting to be employed in the areas of homelessness prevention and alleviation and of risk assessments for survivors of domestic violence. In the latter, survivors are connected with case workers and other relevant service providers with the goal of preventing a recurrence of violence.

These opportunities come with risks, and governments (e.g. Germany) are putting in place a range of measures to ensure they are mitigated, including the appropriate legal, regulatory and accountability frameworks to protect people’s privacy and govern the use of automated systems.

Source: (OECD, 2024[210]; 2025[212]), responses to the questionnaire Harnessing Technology and Data to Improve Social Protection Coverage and Social Service Delivery, administered by the OECD Social Policy Division; (OECD/UNESCO, 2024[131]); (Policy Lab Digital, Work & Society within the German Federal Ministry of Labour and Social Affairs, 2024[213]).

Box 5.50. PES in OECD countries are increasingly turning to AI for service provision

Half of Public Employment Services (PES) in OECD countries have deployed AI-powered solutions, including generative AI solutions using LLMs, to aid their activities and services to jobseekers, people at risk of job loss and employers. As present, they are most used to:

• profile jobseekers and predict a jobseeker’s likelihood of remaining in unemployment

• match jobseekers with suitable vacancies

• aid the design of vacancy postings, including occupational classification

• provide information to clients via chatbots or virtual assistants

• aid the career management and job-search orientation of jobseekers.

For example, since 2024, the Israeli Employment Service has integrated AI engines into its core systems to streamline job matching processes for job seekers. AI solutions offer numerous opportunities for PES, such as improving the targeting of measures and services, optimising data usage, reducing administrative burdens, and enabling more dynamic digital systems. However, PES need to also consider several challenges and risks, including the need to define clear accountability for AI systems, address concerns about transparency and explainability, ensure data quality and privacy and manage potential resistance or skills gaps among both staff and clients. For example, the Austrian PES implemented a ChatGPT-enabled chatbot in early 2024 to provide information on training and career orientation to jobseekers. However, this innovation came under criticism soon after launch, notably due to reports that the chatbot’s recommendations perpetuated differences between men and women.

To address these challenges, a few OECD countries are pioneering innovative steps to promote better governance of AI within their PES. This includes the establishment of multi-disciplinary ethics committees by PES in France and Belgium, and the development of a dedicated strategy by Norway’s PES to set out the ambitions, enablers and principles for AI use.

Source: (Brioscú et al., 2024[214]), https://www.maariv.co.il/business/carrier/article-1122739.

Health services represent a particularly promising field, with AI experts considering it to be one of the areas of the greatest potential for AI (OECD, 2024[29]). AI systems can suggest diagnoses and predict risk factors to help introduce preventative measures (Khalifa and Albadawy, 2024[215]; Alowais et al., 2023[216]). This was particularly helpful during the COVID-19 pandemic. For example, Greece developed and used an ML to optimise border testing by targeting asymptomatic travellers based on demographic data and previous test results (Bastani et al., 2021[217]). In some contexts, AI can also be used to predict the likelihood of illness or injury. For example, in Portugal, the application PrevOcupAI is applied to predict work-related illnesses and connected risk factors in public administration (EC JRC, 2021[218]). In Saudi Arabia, AI is being deployed to predict and reduce patient no-shows at outpatient clinics, significantly improving resource utilisation.6 In Korea, the city of Incheon is using AI technology to improve healthcare accessibility for foreign residents and visitors (Box 5.51).

Combined with the knowledge of doctors and other medical experts, AI can lead to better accuracy, higher efficiency and more positive outcomes for people. The benefits can go beyond citizen well-being; in the United Kingdom, savings to the public budget through expansion of AI-enabled preventive health care are estimated to be worth GBP 6 billion per year by 2040 (TBI, 2024[219]).

Box 5.51. Korea’s AI-powered multilingual healthcare communication

In Korea, the city of Incheon has implemented SIMTOMI, an AI-based medical support service that bridges the language gap between foreign patients and healthcare providers. This service combines a multilingual AI symptom checker with advanced language translation capabilities, allowing foreigners and tourists to accurately communicate their health concerns to local medical professionals and pharmacists. The system generates anonymised symptom summary reports that can be shared via QR codes during in-person hospital or pharmacy visits. The impact of this innovation has been significant, with the service achieving a high satisfaction score of 84.4 points and a 95.0% intention to reuse rate, exceeding initial targets.

Source: https://oecd-opsi.org/innovations/simtomi-ai-powered-multilingual-medical-support.

Forecasting future public service needs

AI also has wide potential uses in conducting predictive analytics to forecast future service needs as well as to prepare and adjust the response capacity accordingly. This can complement the profiling of service users and their needs, helping to better target and propose public services to users. The potential of predictive analytics is well known in the field of policing and justice (Macnish, Wright and Jiya, 2021[220]), as discussed in the “AI in law enforcement and disaster risk management” and “AI in justice administration and access to justice” sections of this chapter. For optimising large-scale systems, AI predictive systems have demonstrated concrete service quality improvements, such as optimising traffic, decreasing road accidents and increasing fuel efficiency in public transportation systems (Jevinger et al., 2023[221]). In Belgium, the Agency for Roads and Traffic used AI is used to predict slippery road conditions and help allocating resources for keeping roads ice free. In the Netherlands, AI is used to experiment with traffic predictions and forecasting road accident. In Portugal, the application PrevOcupAI is used to predict work-related illnesses and connected risk factors in public administration (EC JRC, 2021[218]).

A step further is the emerging trend of proactive service delivery, where AI can help to anticipate people’s needs beyond the current public service context and rapidly respond to them (Pawlowski and Scholta, 2023[222]).7 For example, this may involve citizen digital twin (CDT) systems, which collect data about a person’s overall situation and then helps to structure scaffolding service responses to their life events (Kopponen et al., 2024[211]). In this way, tailored service combinations can be created proactively to meet individuals’ specific needs depending on their current situation or predicted future life event (e.g. finishing school and entering the labour market). In Singapore, the ”Moments of Life” application uses AI algorithms to analyse user information and send personalised notifications (e.g. for upcoming vaccinations or school enrolment deadlines) (Aldemir and Uçma Uysal, 2025[223]).

Box 5.52. How cities are using AI tools to provide services

AI is starting to play a critical role in public service provision and tackling urban development challenges. For example:

• Cities are using AI to optimise traffic and mapping transportation gaps and needs. Los Angeles (United States) uses AI to detect and respond to traffic incidents, such as accidents or road closures, minimising their impact on traffic flow. AI-driven systems analyse data from traffic cameras, sensors and social media to provide real-time alerts and suggest optimal routes to drivers, reducing overall congestion. Sydney (Australia) is using AI to optimise traffic flow and test driverless buses, setting the stage for wider adoption.

• Cities use AI to provide public services and information. Buenos Aires (Argentina) created a chatbot called Boti, first as the official government channel for COVID-19 vaccinations, and now as a channel to provide information on topics such as bike sharing and social care. Yokosuka (Japan) is developing a dementia prevention service using conversational generative AI to strengthen elderly citizens’ mental health.

• Greece’s Ellinikon Project aims to transform the former Athens International Airport into a cutting-edge smart district. Advanced technologies, including IoT, AI analytics, 5G connectivity and ultra-high-speed fibre optics, will be seamlessly woven into the fabric of the city. Smart sensors and sophisticated AI algorithms will be deployed to manage infrastructure efficiently, enhancing resource allocation and service delivery. Smart mobility solutions, such as intelligent parking assistance and shared transportation options, will further streamline urban movement. Additionally, AI will play a pivotal role in optimising water and waste management, ensuring environmental sustainability. A Digital Twin of the city will be created to monitor and optimise resource consumption and environmental conditions in real-time.

Generative AI in particular has the potential transform the way cities work; it can predict demand and streamline public services, improve local governments’ capacity to innovate, enhance data-driven decision-making, optimise resource allocation, and streamline administrative processes. This is possible as some cities collect data continuously from vehicles, buildings, infrastructure and individuals via sensors, drones and mobile devices to present an up-to-the-second picture. For example:

• Amsterdam (Netherlands), Eslöv (Sweden), Helsinki (Finland), and Munich (Germany) use digital twins to improve city planning and management and build more resilient cities to future threats. The digital replica, constantly updated with data from its physical twin, enables cities to analyse data, monitor systems and run simulations and build scenarios each evaluated against key criteria such as social accessibility, environmental sustainability and financial return, facilitating decision-making on housing, land use, density, green infrastructure and transport.

• Shanghai (China) uses GenAI-powered digital twins and IoT to manage its complex city infrastructure by combining real-time data from IoT devices with sophisticated analytics. This helps the city to ease traffic, boost the effectiveness of public transport, optimise energy use, enhance disaster preparedness and plan infrastructure.

Despite its potential, cities are at an early stage of GenAI use. A 2023 Bloomberg Philanthropies survey of 100 cities around the world found that only 2% were actively implementing GenAI in their activities, while 69% were exploring or testing the technology but not yet developing capabilities and policies. Barriers include lack of awareness, budget constraints, a lack of relevant guidelines and data governance, skills gaps, data privacy and security concerns, and administrative and technical challenges in integrating AI technologies such as geospatial AI (GeoAI) in planning and service design processes.

Going forward, to harness the potential and uptake of AI, cities should:

• close digital divides, in coordination with the national government, to ensure access to internet services for the entire population

• consider and proactively address the future challenges, risks and potential drawbacks of AI (e.g. data privacy issues and data quality and availability that limit decision-making)

• develop, with the support of national governments, the regulatory systems to ensure that AI systems used are safe, transparent, traceable, and non-discriminatory to preserve the privacy, ethical values, and individual’s rights

• explore regulatory sandboxes to foster public-private partnerships in a controlled and innovative environment and test ideas safely

• invest in AI literacy for local public employees, develop specialised AI expertise within their legal departments.

• build partnerships with academic institutions and tech companies for knowledge sharing.

Source: (OECD, 2024[224]; 2023[225]; Bloomberg Philanthropies, 2023[226]), https://theellinikon.com.gr/en/smart-and-sustainable/smart-city.

Evidence of impact

In a recent study on the potential automation of UK public services, the Alan Turing Institute found that AI could help automate 84% of service-related transaction and save an equivalent of approximately 1 200 person-years of work every year (Straub et al., 2024[199]). In addition, reduction in red tape and improvements in regulation around public services’ design and delivery could greatly benefit users in terms of saving time, effort and resources, or in accelerating response time from government.

While the promise is great, there is so far little empirical evidence that AI has or will deliver on these benefits. Current knowledge about AI’s adoption in public service is based on case studies that synthesise early evidence from implemented practices.8 Use cases now are multiplying rapidly, and the following sections will give an overview of the emerging innovation trends.

Most of the evidence of impact of the newest AI innovations in public services is based on individual projects and cases and is most often self-reported. Case review in literature indicates a general positive indication of the benefits of AI. However, it is not very common for AI projects to disclose their potential negative impacts in advance (Birhane et al., 2022[227]). This is expected, as the applications are still relatively new, issues are still emerging, and evaluations are often not carried out.

As AI continues to reshape public services, rigorous testing of behavioural interventions is essential to ensure these technologies are not only efficient but also effective in influencing citizen engagement and decision-making (Soman, 2024[228]). Some government teams are already prototyping and rapidly testing AI-driven solutions. However, more structured evaluation methods — such as RCTs, difference-in-differences analysis, and A/B testing — can help determine which AI applications work best, for whom, and in which contexts. Partnering with academics to make these experimental approaches part of AI implementation can provide stronger empirical evidence on AI’s real-world impact, ensuring public services are designed and adapted based on behavioural insights rather than assumptions.

Managing risks and challenges

While AI has tremendous potential to improve the efficiency, effectiveness and responsiveness, there are risks and challenges that can limit adoption and lead to real-world harms. A 2022 study on deriving lessons learned from 61 cancelled public service automation initiatives across various countries finds that many have failed due to lack of effectiveness, civil society protect, critical media investigation and legal action, among other issues (Redden et al., 2022[229]). The top risks and challenges identified are discussed below.

Associated risks

• Misuse or questionable use of AI, resulting in surveillance and privacy concerns

• Inadequate or skewed data in AI systems

• Lack of transparency and explainability

• Exacerbating social exclusion and digital divides

Using AI in public service innovations requires increased access to users’ personal information, especially when connected to proactive and personalised service delivery. Often this also requires having access to a variety of public organisation databases and service information to identify to best service approach. Proactive service delivery, for example, often requires algorithmic processing of data scattered across multiple publics and sometimes private data sources (Nikiforova et al., 2023[230]). The use of AI raises concerns about privacy and the purposeful and responsible use of data and can make these services more contestable (see Chapter 1, section on “To seize the benefits of AI, its risks need to be managed”). A commonly held assumption is that intelligent and data-trained tailored AI tools face inescapable trade-offs between (predictive) capability and privacy. However, there has been much recent research interest in getting around this general hypothesis thanks to privacy preserving technologies (PPTs) or privacy enhancing technologies (PETs), which use anonymised or encrypted data (OECD, 2023[231]).9

If AI systems rely upon inadequate or skewed data, it could lead to inaccurate or adverse outcomes for some individuals or groups. With regard to public service design and delivery, this could result in, for instance, undue preferential treatment for some individuals or groups or reinforcing inaccurate stereotypes. Related in part to inadequate or skewed data, digital data divides — where some groups are more represented in data than others — limit the potential for personalised AI services, leaving them only useful and accurate for data-rich populations and potentially leaving some groups less represented (UNESCO, 2019[232]; Perry and Lee, 2019[233]; Dieterle, Dede and Walker, 2024[234]).

Given the nature of public services, ensuring the accountability, transparency and explainability for decision-making and results from AI systems and tools is crucial. This requires the creation and implementation of public service frameworks that define clear lines of responsibility for AI systems along their entire lifecycle within public services. The more transparency there is in the use of AI in public services, the easier it will be to monitor for adverse effects and create a greater sense of explainability and trust in AI outcomes (Sun and Medaglia, 2019[197]; Thiebes, Lins and Sunyaev, 2020[235]).

Implementation challenges

• High costs of AI adoption and scaling

• Lack of actionable frameworks and guidance on AI usage

• Difficulty in establishing continuous monitoring for track progress and surface risks

• Lack of high-quality data and the ability to share it

• Skills gaps

The cost associated with building and maintaining AI systems for public services (from infrastructure, monitoring, staff training and education) cannot be ignored. For instance, the cost of commercial AI subscription options could increase when companies seek to become profitable and cover large overhead and operating expenses to satisfy investor expectations (Shark, 2025[236]).

The trustworthy adoption of AI systems in the design and delivery of public services requires continuous monitoring to track progress and help identify emerging risks, but governments generally lack this function. Fully autonomous AI systems that have an impact on citizens’ lives can present legal issues and further contribute to distrust and resistance in adoption. As a result, human determination and oversight is often required or otherwise prioritised in public policy. For instance, the EU AI Act imposes human oversight requirements on “high-risk” use cases, which many government uses are considered. However, relying on human intervention alone may create a false sense of security in adopting algorithms (Green, 2022[237]). Having humans in the loop does not negate the need to carry out cost-benefit and benefit-risk analyses to weigh trade-offs, nor for evaluation throughout the AI lifecycle. Consequently, the adoption of AI in public service provision needs granular and technical evaluation from a public value perspective in a where adverse effects may manifest during AI system implementation (Medaglia, Gil-Garcia and Pardo, 2021[238]; Madan and Ashok, 2023[194]).

Some of the early failures in AI use in public services and instances of errors have been connected to data quality issues. OECD (2019[191]) research has identified to the need to continually verify the accuracy, reliability and appropriateness of data inputs to AI systems, which helps service providers to identify and address errors present in the data.

Lastly, varying levels of AI competences and associated skillsets in government are problematic (Medaglia, Mikalef and Tangi, 2025[239]). While generative AI tools have become widely accessible for bottom-up adaption, there has not been sufficient training in AI in government and clear guidelines for its use in public organisations for developing and informing public services. Furthermore, given the rapid development of AI systems continuous learning is a must to keep current. Necessary skills for adapting AI systems are not only technological but also ethical, legal, managerial and sociotechnical; public administrators need to understand how the AI systems interact with both the organisation and the public at large (Trajkovski, 2024[240]). So, for example, creativity and design thinking skills are necessary to implemented human-centric services tailored to public needs. This also requires high levels of interpersonal skills, as administrators need to effectively collaborate with data scientists, lawyers, ethicists and external partners. A lack of these skillsets may cause delays in AI adoption or worse negative impacts, such as biased or opaque algorithmic systems. Systemic approaches to developing such skillsets across governments are currently largely lacking. Some have proposed a new profession of AI functional professionals who act as intermediaries between the technical and public policy side and help train public officials accordingly (Labanava et al., 2022[241]).

Untapped potential and way forward

There is significant potential to using AI in additional service areas and in novel use cases. For instance, in healthcare, AI could enable tailored and preventative interventions and informed behavioural “nudges” leading to better outcomes and allowing health professionals more time for care. It could also yield new techniques to unlock value from vast health data assets, 97% of which remain untapped in OECD countries (Sumner et al., 2023[242]; Bennett Institute, 2024[243]; OECD, 2024[244]). In this area, the OECD (2017[245]) standard for promoting the use of health data in the public interest guides governments in making this data more available for innovation in health care. Exploration of challenges for the implementation of the principles of this recommendation may be found in (OECD, 2022[246]) and in a forthcoming OECD report: “Facilitating the Secondary Use of Health Data for public Interest Purposes Across Borders”. AI advances could also help ease projected healthcare workforce shortages of 3.5 million by 2030 in OECD countries (OECD, 2024[244]).

AI use in public service design and delivery is still in an experimental phase. While there are examples of automation, streamlining and even new types of services emerging, government organisations are still far from broad adoption and operational use. AI hype and demands for cost efficiency may push governments towards a rapid automation. Yet evaluations and critiques of early practices, as discussed in this section, suggest that a more gradual approach taken together with public servants may be more likely to succeed. Governments need strategies on how to use hybrid models where AI augments human decision-making, and on how to deal with the evolving role of public officials in this new format of algorithmic bureaucracy (Vogl et al., 2019[247]). While the promise of AI may be to free up public servants’ time to higher-value tasks — such as engaging with citizens to tackle complex social issues — it can also have the opposite effect, distancing public servants from society as their role in direct data collection/interaction with service users diminishes (Bullock, Young and Wang, 2020[248]; Madan and Ashok, 2023[194]) (see also the discussion on “AI in civil service reform” in this chapter).

Another emerging approach to automation in public services is rules as code (RaC), which involves encoding government rules into machine-consumable formats alongside traditional legal text (Mohun and Roberts, 2020[4]). As services increasingly rely on AI to support decision-making, RaC provides a foundation for more accurate, transparent and scalable AI applications in government. By embedding rules in digital formats, RaC could enhance the consistency of AI-powered service delivery, ensuring that automated systems interpret and apply regulations correctly across different contexts. Several governments have begun using RaC to streamline regulatory compliance and enhance AI-driven service delivery. For example, the United Arab Emirates has launched a nationwide Rules as Code platform that aims to develop AI-based laws and regulations, transforming the financial ecosystem.10

Unlike conventional task-based approaches — where service delivery is segmented into distinct sequential operations — AI-enabled approaches can integrate functions and decision-making processes simultaneously. This represents a potentially transformative shift in how public services are organised.

When it comes to the broader acceptance of AI-based public services by citizens and residents, a few clear insights emerge from existing research. It is generally accepted that AI is better at processing large amounts of data and identifying patterns (Rane, Choudhary and Rane, 2024[249]). However, a combination of factors including perceived usefulness of AI systems, performance expectancy, attitudes, trust and effort expectancy tend to influence the willingness to use of AI across sectors (Kelly, Kaye and Oviedo-Trespalacios, 2023[250]). Perception of risk and trust in AI systems differs when dealing with general public services (i.e. provided by the government without specific request, and concern all or the majority of citizens, such as basic education and public safety) versus specific public services (i.e. explicitly requested by citizens and impact only one or a few citizens, such as elderly care programmes or housing assistance for low-income populations). General services are more abstract and the users have lower levels of situational awareness meaning that they may also accept AI more easily.11 In specific services, the opportunity to decide becomes relevant; users general want to have a choice whether an AI application has been used in their service decision (Gesk and Leyer, 2022[203]).

Early research also indicates that trust in AI applications is higher if they are integrated into already existing services rather than services that are completely new (Aoki, 2020[251]). This means that more proactive service innovations may have to prove themselves to users and may face more scrutiny. In all cases, participation of users in co-creating public services is correlated to a positive perception of AI decisions and could drive higher adoption (Gesk and Leyer, 2022[203]). This is very important, as people’s experiences with services can influence how they perceive their governments overall, and public satisfaction with administrative and social services is an important driver of trust (OECD, 2024[46]). Public organisations can risk their democratic legitimacy if the public does not trust the services government intend to provide with AI (Aoki, Tay and Yarime, 2024[252]; Aoki, 2020[251]).

Governments also need to better understand and consider user needs to design AI-enabled services they are likely to want to use. This will involve ensuring that policies and services are designed for how people actually behave, rather than how they are assumed to behave. Even well-intended policies can fail when they introduce unnecessary friction ("sludge"), making it harder for people to access services, complete applications or make informed decisions. Further use of behavioural science can assist this by offering practical levers to reduce sludge, improve accessibility and ensure services are user-centred and trusted by citizens. By addressing cognitive barriers — such as complexity, decision fatigue and inertia —governments can design services that are simpler, more intuitive and easier to navigate. AI-powered tools can identify services that need sludge audits, assess intervention effectiveness and improve service design. NLP and sentiment analysis can analyse public feedback in real time, indicating pain points across online platforms and customer service channels. AI can also quantify user interactions and tailor interventions to different demographics, ensuring policies are better targeted and more inclusive. As governments move towards digital-by-design and data-driven governance, integrating behavioural insights with AI will be key to making public services more user-friendly, efficient and responsive (OECD, 2024[253]).

Finally, there is a considerable gap between the speed at which AI is being introduced to public services and the extent to which robust evaluations are carried out. The more thorough analyses already point to many unforeseen and sometimes adverse effects that may appear next to general efficiency gains. This area needs further investment.

In all of these areas, international cooperation in building public sector AI capacity, is crucial. The more government share AI practices across governments in public service development (including sharing open algorithms, infrastructure, intergovernmental datasets, and joint efforts for the responsible development of emerging technologies), the more likely that quality across the board can be assured. Multilateral initiatives such as the Global Partnership on Artificial Intelligence (GPAI) can play a crucial role in this regard.

Notes

← 1.  https://oecd-opsi.org/innovations/use-of-robotic-process-automation-in-forensic-routines/

← 2.  https://oecd-opsi.org/innovations/streamlined-benefits-for-a-more-equal-society/

← 3.  See https://cms.hacienda.cl/cmod_cl/assets/documento/descargar/3a29739655166/1712674476.

← 4.  See https://ai.gov.uk/projects/caddy.

← 5.  https://www.tech.gov.sg/gtpc-product-1.

← 6.  https://oecd-opsi.org/innovations/no-show-prediction-at-outpatient-clinic.

← 7.  Proactive public services are services that a public organisation “pushes” toward its counterparts based on their “needs, circumstance, personal preferences, life events and location” (Linders, Liao and Wang, 2018[321]).

← 8.  Use cases for AI have centred around high-stakes, data-rich areas including safety and security (e.g. risk detection, prediction and simulation (Margetts and Dorobantu, 2019[322]; Xavier et al., 2022[323]; Zhang, Atasoy and Vasarhelyi, 2022[324]), facial-recognition in policing (Zilka, Sargeant and Weller, 2022[325]), recidivism prediction in criminal justice (Kleinberg et al., 2017[326]), process automation (Ojo, Mellouli and Ahmadi Zeleti, 2019[327]) and forecasting future needs in services (Bright et al., 2019[328]).

← 9.  See https://www.oecd.org/en/topics/privacy-enhancing-technologies.html.

← 10.  https://u.ae/en/about-the-uae/digital-uae/digital-technology/artificial-intelligence/rules-as-code.

← 11.  First, public services can be divided into general (generally provided, e.g. traffic lights) and specific (individually requested and specifically directed).