This paper reviews how governments are turning responsible AI principles into practice, offering practical guidance and examples to help operationalize AI.
This EFI Insight distills critical insights from the implementation of Beneficial Ownership Registers (BORs) in Nigeria, North Macedonia, Kenya, and the United Kingdom. The experiences of these countries offer valuable lessons for similar reform efforts worldwide aimed at enhancing beneficial ownership transparency.
This guidance note provides useful advice to policymakers on how to implement coordinated GovTech and Green policies, underlining the importance of adequate leadership and commitment.
The new GTMI data dashboard helps users visualize and explore the contents of the GovTech Dataset and the GovTech Projects Database.
mGovt offers many opportunities to connect citizens to their governments, but challenges remain. This presentation will feature the latest mGov How-to-Note, which offers strategies for maximizing opportunities and overcoming implementation challenges.
Digital technologies and data have the potential to increase substantially the tools that governments have available to further involve citizens in policy and service design and delivery. This how-to-note provides advice on how to use digitalization to strengthen the engagement between the governments and citizens, with various examples of what CivicTech is and why it is an important element of the GovTech approach.
The country examples below illustrate how AI is already being applied in governance settings across the world. To explore further and contribute examples from your own country, visit the World Bank AI Repository — a platform developed in collaboration with multilateral development banks that invites governments, practitioners, and researchers to share real-world applications and help build a global knowledge base for responsible AI in development.
Albania's EU accession process demands the systematic alignment of its national legislation with the EU acquis — a vast, technically complex, and continuously evolving body of European law. Doing this manually is slow, resource-intensive, and prone to inconsistency, creating a bottleneck that constrains the pace and quality of European integration. To address this, the National Agency of Information Society integrated an AI-driven harmonization tool into the European Integration and Membership Platform (PIEA). Built on GPT-4.0 and advanced Natural Language Processing (NLP), the system automates four core tasks: it translates EU legal acts into Albanian with over 95% reported accuracy; it performs comparative legal analysis to detect overlaps, gaps, and contradictions between EU and Albanian law; it drafts and aligns transposed legal acts with national legislation; and it auto-generates compliance tables for documentation and monitoring. The primary data inputs are EU legal texts, Albanian national legislation, and regulatory compliance documentation. The technology was co-developed with the private sector and academia. Key outcomes to date are primarily capacity-building in nature: 421 members of the Inter-Institutional Working Group for European Integration have been trained, out of 656 invited. The most significant challenge has been ensuring data quality — variations in legal language, document structure, and translation quality across both EU and Albanian source documents can affect the accuracy of comparative outputs. The tool is fully operational and represents one of the more advanced deployments of generative AI in legislative harmonization globally.
A persistent gap in Argentina's judicial response to gender-based violence has been the absence of structured, unified data on how such cases are handled by the courts. Without reliable data, it is difficult to identify systemic patterns, design evidence-based prevention strategies, or hold the justice system accountable. AymurAI, developed by Fundar in collaboration with Criminal, Misdemeanor and Administrative Offenses Court No. 10 of Buenos Aires, addresses this by using a semi-automated AI approach to extract and classify key information from unstructured judicial rulings. The tool uses Named Entity Recognition (NER) models to identify structured data from scanned PDFs and digitized legal documents — including the type of violence, the relationship between parties, frequency of incidents, and court decisions — while incorporating human validation to ensure accuracy and reliability. The outputs are published as open datasets, enabling policymakers and civil society to analyze trends and design better interventions. The initiative is not primarily generative AI, relying instead on classification and extraction models. Key results include the publication of 6,569 court rulings, of which 30% concerned gender-based violence. Within those cases, 54% involved physical violence, 78% involved partners or family members, and 64% of incidents were recurrent. The biggest challenge was data quality: judicial records existed in inconsistent, unstructured formats that required significant preprocessing. The open-source nature of AymurAI has allowed the methodology to be replicated, with early interest from other Argentine provinces.
Brazil's judiciary faces a well-documented challenge of predatory or "abusive" litigation — the mass filing of repetitive, poorly substantiated lawsuits by professional litigants or law firms using automated systems, often without individualized legal arguments or supporting documentation. This floods civil and small claims courts, delays legitimate cases, and disproportionately harms vulnerable citizens. Developed by the Innovation Laboratory of the Pernambuco Court of Justice (IDEIAS/TJPE), Bastião is an AI-powered tool that analyzes large volumes of court filings to detect patterns indicative of abusive litigation: repeated parties, identical legal theses, missing individualized claims, and anomalous filing volumes. Using NLP and pattern recognition on case metadata and legal texts extracted from official case management systems, the tool generates visual and textual analytical reports that help judges and clerks prioritize caseloads and flag suspicious clusters for further review. Bastião was built entirely in-house using agile methods and existing infrastructure, making it low-cost and potentially replicable. Deployed in October 2023, the tool has improved procedural efficiency, enabled court staff to redirect time to more complex and legitimate cases, and sparked institutional conversations about the ethical use of AI in judicial settings. All data processing complies with Brazil's General Data Protection Law (LGPD). The most significant challenge was institutional resistance: initial concerns among judges and staff about whether AI could reliably identify abusive behavior, and fears that it might undermine judicial autonomy. These were addressed through ongoing engagement, demonstration pilots, and clear communication that Bastião supports — rather than replaces — human judgment.
Brazilian courts face a chronic backlog of cases, and a significant share of the daily workload of judges and clerks consists of repetitive drafting tasks — writing decisions, orders, and reports for standard case types. The Biblioteca de Prompt Jurídico (BPJ) addresses this through a deceptively simple but effective innovation: a structured, searchable digital library of well-crafted prompts designed for use with AI tools such as ChatGPT, created by a judge based on real courtroom experience. The library organizes prompts by type of judicial process, legal area, case class, and type of judicial act, and integrates a filtering system aligned with the PJe (Processo Judicial Eletrônico) classification system, so users can quickly find the right prompt for a specific task. The system does not replace judicial decision-making — the judge still reasons and decides — but dramatically reduces the time needed to draft the output of that reasoning. A striking example: a decision on a motion for clarification (embargos de declaração) that previously took two hours can now be drafted in 43 seconds using an appropriate prompt. The BPJ relies on classification data from the PJe system and textual data from judicial practice. It was developed independently by a judge, not as an institutional initiative, and operates with minimal infrastructure. The primary remaining challenge is expanding the collection — building a larger, more diverse prompt library covering a wider range of judicial situations requires collaborators that the initiative currently lacks.
Ethiopia's judicial system has historically struggled with case backlogs, inefficiencies in documentation, and significant accessibility gaps for citizens — particularly in rural areas and for those dealing with local-language proceedings. The Smart Court System, developed by the Ethiopian Artificial Intelligence Institute (EAII) in collaboration with the Federal Supreme Court, provides a comprehensive digital platform to address these challenges. Its key components include a Smart Digital Information Desk that allows citizens and lawyers to track cases and appointments; an AI-powered chatbot integrated with Telegram for filing complaints and accessing case information; a Digital Audiovisual Recording system that captures live court proceedings; an AI-driven audio transcription tool capable of converting spoken proceedings in local Ethiopian languages into written records; and support for online testimonies and virtual court sessions. Primary data inputs are legal texts, court documents, and audio-visual recordings. The system was developed fully in-house by government staff. Reported outcomes include a reduction in case backlog, faster documentation through automated transcription, improved accuracy and accessibility of legal records, and reduced operational costs. The most significant challenges are twofold: limited availability of high-quality training data in Ethiopia's diverse local languages, which constrains the scalability of AI transcription across regions, and securing adequate funding for ongoing development and deployment.
India's judiciary handles one of the largest volumes of litigation in the world, with chronic backlogs, lengthy resolution times, and significant inefficiencies in document management and case processing. Nyaay AI is an end-to-end AI platform designed to automate and streamline judicial workflows, serving courts, law firms, and legal institutions. Its core capabilities include: automated content extraction from case documents; summarization of lengthy legal texts; classification and tagging of case types; prediction of likely outcomes and case durations based on historical data; linkage of related cases to surface precedents; e-filing automation that reduces redundant data entry; judgment analysis and headnote generation; and translation of judgments and legal documents between languages to improve accessibility. The platform uses a combination of NLP, machine learning, and proprietary algorithms, adapted from open-source frameworks, with legal documents and historical case data as its primary inputs. Nyaay AI is operational with the Indian Supreme Court, various High Courts, and the Singapore Supreme Court, as well as commercial law firms. In collaboration with BITS Law School, Nyaay AI co-founded the PALETTE Centre — India's first dedicated hub for AI-powered legal education — inaugurated by former Chief Justice Dr. D.Y. Chandrachud. The platform supports on-premise deployment for data security and compliance. The biggest challenges have been institutional resistance within a traditionally manual legal system, and concerns around data privacy and regulatory compliance — both of which are described as "now changing" as trust in the technology grows.
In Ukraine, the possession or distribution of pornographic material — including private messaging — is criminalized under Article 301 of the Criminal Code, resulting in hundreds of criminal convictions per year for content that many legal experts and civil society advocates consider non-harmful. Court rulings in these cases frequently reference deprecated laws or international treaties Ukraine is not party to, highlighting inconsistencies in legal application. The Better Regulation Delivery Office developed "301 Monitor" as an automated tool that scrapes publicly available court rulings related to Article 301 from Ukraine's official court registry, generates AI-produced summaries of each ruling using generative AI (LLMs), and automatically posts these to a public Telegram channel. The primary data inputs are court ruling metadata and the full text of court decisions. The technology was developed by a private sector vendor. The initiative's goal is to build public awareness and political support for a decriminalization draft law currently before parliament, by making the scale and nature of these convictions visible and accessible to journalists, activists, and the public. The channel is functioning as intended, with news of new convictions spreading through social media and stimulating public debate on decriminalization. No significant technical challenges were reported.
The regulatory approval of nuclear-related shipments in the UAE has historically required intensive manual review processes, creating bottlenecks, delays, and inconsistency in decision-making. NuTech AI is an intelligent shipment assessment system co-developed with the private sector that assigns AI-generated risk scores to incoming shipments based on pre-configured criteria drawn from legal texts and sensor data. Shipments that fall within acceptable risk thresholds are automatically cleared, while high-risk cases are flagged for human review. The system uses NLP and predictive classification models to analyze shipment data and apply risk frameworks consistently. It is currently in a pilot stage. Expected outcomes include a significant reduction in manual workload, elimination of processing bottlenecks, faster turnaround times, improved regulatory compliance, and better allocation of expert human resources to genuinely complex or high-risk cases. The most significant challenge cited is funding for further development and scaling.
Access to legal information in the UAE — and by extension, compliance with UAE law for individuals and businesses — is constrained by the complexity, volume, and technical language of the country's legislative corpus. The Ministry of Justice's Virtual Legal Consultant AI aims to address this by providing an AI-powered interface through which citizens, legal professionals, and businesses can query UAE law and receive precise, sourced responses. The system uses Retrieval-Augmented Generation (RAG) combined with semantic embeddings, allowing it to retrieve and cite relevant legislation accurately rather than relying on model memorization alone. The UAE Legislative Database — containing over 4,993 pieces of legislation and 56,848 legal articles dating from 1971 to the present — serves as the exclusive data source, ensuring responses remain grounded in verified legal texts. The system is hosted on FedNet Azure cloud for security and scalability. While still at concept stage, expected impacts include reduced cost of legal research, improved business compliance, faster legal proceedings, and enhanced public access to legal information. Challenges include the complexity of Arabic legal language (requiring specialized AI models), regulatory compliance constraints, public adoption, and scalability to cover local as well as federal laws and rulings.
Online sexual exploitation of children and youth — including grooming and luring through digital platforms — is a fast-evolving threat that is difficult to detect at scale through human monitoring alone. The Government of South Korea developed an AI-based Proactive Response System to identify warning signs of online grooming and luring behavior across digital channels. Using text mining and data mining techniques, the system automatically classifies collected text to detect patterns associated with advanced forms of sexual exploitation of minors. Once suspicious content is flagged, it can be automatically blocked or reported 24 hours a day, seven days a week, without waiting for human review — enabling a shift from reactive investigation to proactive, technology-enabled prevention. Primary data inputs are text-based digital communications and online content. The initiative allows authorities to intervene earlier in the exploitation lifecycle, with the potential to prevent harm before it escalates to physical contact or more serious offenses.
The Gwangju Metropolitan Police Agency handles approximately 30,000 sexual crime cases annually, yet operates with limited personnel — creating serious capacity constraints in the quality and speed of investigations and victim support. To address this, the agency implemented an AI-powered Sexual Crime Crisis Response Platform using hyper-scale large language models (LLMs). The system performs three core investigative support functions: it summarizes victim statements — which can run to many pages — into concise and accurate briefings; it classifies the type of sexual crime based on case details; and it surfaces similar past cases and relevant legal information to support investigators in building cases and applying law correctly. The system also generates administrative support reports, reducing documentation burden on officers and allowing more time for direct casework and victim engagement. Primary data inputs are case filings, victim and suspect statements, crime classification records, and legal precedent databases. The initiative demonstrates how AI can expand effective investigative capacity without additional headcount, while improving consistency in the handling of sensitive, high-volume criminal cases.
Citizens and businesses engaged in customs disputes — increasingly common due to the growth of overseas direct purchases — have historically had to navigate complex legal databases using keyword searches to find relevant statutes and precedents, a process ill-suited to non-specialists. The Korea Customs Service developed an AI-powered chatbot that allows users to search customs dispute precedents and regulations using natural language queries. The system uses Retrieval-Augmented Generation (RAG) to identify and retrieve relevant statutes and precedents from legal databases, and automatically organizes the factual circumstances and legal reasoning applicable to a user's situation — shifting from rigid keyword-based retrieval to a conversational, context-aware legal research experience. Primary data inputs are legal statutes, customs regulations, and a corpus of customs dispute precedents. The tool reduces the time and expertise required to navigate customs law, supports more consistent application of regulations, and improves accessibility of legal information for a growing population of individual importers and traders.
Employers in South Korea are required to establish employment rules — internal workplace regulations governing conditions of employment — and these must comply with applicable labor law. Labor inspectors have traditionally had to manually review submitted employment rules against a complex and evolving body of legislation to determine compliance, a process that is time-intensive and susceptible to inconsistency across the large volume of employer submissions received. An AI system was developed to automate this compliance determination, reading drafted employment rules submitted by workplaces and assessing whether their provisions conform to current legal standards. The system flags non-compliant provisions and indicates the relevant legal standard being violated, supporting inspectors in their review and enabling more systematic, consistent compliance determinations. Primary data inputs are workplace employment rule documents and the current body of labor law and regulations. By automating routine compliance checks, the system allows inspectors to focus on complex or disputed determinations and helps ensure more uniform enforcement of labor standards across employers.
Corruption in public institutions is often underreported due to fear of retaliation, lack of accessible reporting mechanisms, and low public trust in anti-corruption agencies. Ethiopia's National Corruption Reporting System (NCRS) addresses this by providing a secure, user-friendly platform through which citizens and officials can anonymously report corruption incidents. The system collects, organizes, and analyzes reports using AI algorithms to detect patterns across cases, prioritize investigations, and generate automated alerts for relevant agencies. For public institutions, the NCRS provides case management dashboards and monitoring tools to improve responsiveness and track outcomes. Primary data inputs are citizen reports and legal case records; the system was built entirely in-house by government staff. Reported outcomes include an increase in the volume and quality of actionable corruption reports, faster response times enabled by AI-driven case prioritization, and improved trust between the public and anti-corruption agencies. The single biggest challenge is a dual one: ensuring robust data security and user anonymity to sustain citizen trust and encourage reporting, while also securing sustainable funding for ongoing system maintenance and improvements.
While high-profile corruption cases in Ukraine receive media attention, hundreds of lower-profile criminal court proceedings for corruption occur every month, largely invisible to ordinary citizens. This opacity erodes public trust in government institutions and makes it difficult to assess the true scale and nature of corruption across the country. The Better Regulation Delivery Office developed Corruption Radar — an automated pipeline that scrapes publicly available proceedings from criminal courts for corruption-related cases, uses generative AI (LLMs) to produce short, accessible summaries of each verdict, and automatically posts these to a public Telegram channel (). The primary data inputs are structured court verdict metadata (date, court name, hyperlink) and the unstructured full text of verdicts. The technology was provided by a private sector vendor. Since launch, the channel has gained over 100 followers, primarily journalists, who use the summaries as a source for reporting. The initiative is a low-cost, high-transparency tool that transforms dense legal text into accessible public information, effectively crowdsourcing accountability monitoring. No significant technical challenges were reported.
A recurring dysfunction in large public administrations is organizational fragmentation — where multiple government units share overlapping mandates, leading to duplication of effort, policy incoherence, and coordination failures. Addressing this requires a diagnostic methodology capable of processing large volumes of legal and institutional texts. This Argentine pilot developed a novel methodology using big data and NLP to algorithmically detect similarities in the mandates of public sector organizations based on legal documents such as decrees and laws. Through clustering and text analysis techniques applied to open-source tools, the initiative maps areas of overlap and fragmentation across agencies, providing diagnostic insights for redesigning government structures or improving inter-agency coordination. Key data inputs are legal texts and organizational mandate documents. The methodology enabled the first large-scale, data-driven diagnosis of this kind in Argentina, and was subsequently adapted and applied by the Government of Paraguay. The primary ongoing challenge is institutional capacity: the lack of technical skills and resources within government to sustain, maintain, and operationalize the methodology into routine decision-making represents a significant constraint on its long-term impact.
Government audits are essential oversight mechanisms, but auditors face an ever-growing volume of financial data, procurement documents, contracts, and official records that are difficult to process manually with the speed and consistency required. The Lianyungang City Audit Bureau in Jiangsu Province, China, developed what is described as the first public government audit application of its kind in China, built on Baidu's AI platform. The tool assists auditors in three primary ways: it helps identify the correct legal characterization of findings during audit work; it provides audit strategies, methods, and recommendations; and it enables analysis of financial and procurement data at scale. Data inputs include financial records, business data, official documents, procurement and bidding documents, and contracts. The system uses NLP for text analysis and predictive classification for pattern recognition. The tool is fully operational and was co-developed using Baidu's AI platform (appbuilder). Challenges noted include selecting the right large language model and ensuring appropriate reference to relevant regulations. The initiative reflects a broader trend in China toward AI-assisted governance and is presented as a model for intelligent government governance that can deliver more efficient and equitable oversight.
Real estate transactions in Georgia have historically required the physical presence of parties, involvement of intermediaries such as notaries, lawyers, and brokers, and manual identity verification — all of which add time, cost, and friction to property registration. The NAPR's Smart Contract initiative digitizes the full cycle of real estate transactions, enabling Georgian citizens — both within the country and abroad — to remotely execute sales, purchases, and mortgage transactions in a secure digital environment. The process covers four stages: identification of parties, agreement on contract terms, transfer of funds, and registration of title changes. AI is used specifically for Remote Identity Verification, comparing users' personal identification documents and photographs against the Public Service Development Agency's database. This is supplemented by video and audio recordings of transactions, which are archived securely in NAPR datacenters. The system integrates flexible payment options through commercial banks, including escrow services. Built entirely in-house, the tool is fully operational and has already facilitated both cross-border and domestic real estate transactions. The primary challenge has been resistance to change from the public and from public sector employees accustomed to traditional processes — a challenge that the institution anticipates will diminish as users experience the full benefits of the digital service.
Documenting government workflows and business processes is a necessary but labor-intensive activity that often requires skilled analysts and significant time investment. The Federal Tax Authority is implementing an AI-powered Process Capture tool (Nintex Process Manager/Promapp) that records real-time user interactions — clicks, keystrokes, and screen activity — and automatically converts these into structured, step-by-step process maps and editable documentation using generative AI. The system allows users to review, edit, and finalize procedures with minimal manual effort, and is designed to enhance accuracy, consistency, and speed in process documentation, while supporting faster onboarding and institutional knowledge transfer. Primary data inputs are user interaction data (screen activity, keystrokes, application usage) and process metadata (system names, timestamps, step descriptions). Expected outcomes include enhanced process mapping efficiency, reduced manual effort, and improved documentation accuracy. The initiative is still being developed. No major challenges have been encountered, though security policy requirements have necessitated additional confirmation from the vendor (Nintex) before full deployment.
Allocating work permit quotas to over 600,000 business establishments in the UAE has historically been a manual, rule-based, and committee-dependent process with turnaround times of approximately four days per request, high error rates, and significant administrative burden. MOHRE's AI Quota Allocation System modernizes this through a dual-model approach: for new companies, a K-Nearest Neighbors (KNN) model assigns quotas by comparing new establishments to similar existing ones based on demographic and sector features; for existing companies, ensemble models (Random Forest and XGBoost) predict quota needs based on historical patterns, improving accuracy and reducing unnecessary transactions. The system is integrated into MOHRE workflows via real-time API (ESB), enabling instant, proactive quota decisions without committee review except for exceptional cases. Since January 2025, the system has achieved approximately 2x better accuracy compared to the manual process, reduced recurring transactions by 3x, cut committee workload by roughly 75%, and achieved an 88% success rate for new-company quota allocations (compared to 49% under manual rules). A 67% reduction in manual review efforts for additional quota requests has been reported. The primary challenges were the complexity of integrating data from multiple systems (processing millions of historical records for model training), and ensuring that the AI system allocates quotas fairly and transparently across sectors and nationalities while maintaining compliance with labor market policy and avoiding bias.
Labor inspectors in South Korea handling wage disputes, workplace violations, and labor law infractions have traditionally had to manually review extensive documentation — including lengthy witness statements from both complainants and suspects — before being able to draft petitions or identify the core legal issues in a case. The Ministry of Employment and Labor's AI Labor Inspector Support System deploys large language model (LLM)-based tools — drawing on models including ChatGPT, Llama, and MAAL — to function as an "AI colleague with 30 years of experience" for inspectors. The system automatically summarizes witness statements spanning dozens of pages, identifies key issues in disputes such as unpaid wages, drafts petition documents, and surfaces relevant laws and legal precedents. A parallel citizen-facing service accessible via Kakaotalk provides 24/7 labor law consultation for workers, serving over 2,500 citizens with nearly 12,000 inquiries in its first month alone. Primary data inputs are legal statutes, official Q&A records, case precedents, and witness statement texts. The initiative reduces the time burden on inspectors, improves the consistency and quality of legal reasoning in labor dispute processing, and extends access to legal guidance for workers outside of office hours.
Determining the correct industry classification of a workplace is a prerequisite for setting accurate insurance premium rates, yet the process has historically required manual review by trained specialists and could take up to two months per case, creating significant administrative backlogs. An AI-based OCR and recommendation model was developed to automate this classification task, reading business-related documentation and recommending the correct industry category. The system achieves 92% accuracy and reduced processing time from a maximum of two months to approximately 30 seconds — a compression of several orders of magnitude. Specialist staff are freed to focus on complex or disputed cases that genuinely require human judgment, while routine determinations are handled automatically. Primary data inputs are business registration documents and workplace-related documentation. The initiative demonstrates the potential of AI-driven classification systems to substantially reduce administrative backlogs in regulatory functions that depend on large volumes of structured determinations.
The Korea Workers' Compensation & Welfare Service processes over 160,000 compensation cases annually, each requiring a determination of eligibility based on the nature of the injury or illness, relevant precedents, and prior disease determination results. The volume and complexity of this workload creates significant pressure on decision-making staff and risks inconsistency across determinations. An AI system was developed to search across the full corpus of historical cases and precedents, training on past insurance decisions and disease determination outcomes to support and streamline case officer decision-making. The tool surfaces relevant precedents for each new case, reducing the time required for legal research and improving the reliability and consistency of decisions across a high-volume caseload. Primary data inputs are historical case records, insurance decision judgments, review decisions, and medical determination precedents. The initiative improves throughput and decision quality in an administrative function with direct consequences for injured and ill workers.
Tax evasion and fraud represent a major fiscal challenge for Armenia, compounded by manual, unfocused risk assessment processes that limit the effectiveness of audit targeting and revenue collection. The State Revenue Committee's Tax AI initiative uses machine learning — including both structured and unstructured learning approaches — to identify anomalies, detect fraud, and flag potential tax evasion cases from a range of data inputs including receipts, tax declarations, historical tax returns, employee statistics, and financial records. The system also supports risk management and audit planning by generating AI-driven insights and incorporates generative AI elements to produce text summaries and draft reports. The technology was co-developed with private sector and academic partners. The primary reported outcome is an AI algorithm that exceeds 85% accuracy in anomaly detection at the time of submission, with accuracy continuing to improve. The initiative has been recognized internationally, including in an IMF PFM Blog post, as a leading example of AI-driven tax administration. Key challenges include data quality, limited technical skills within the organization, and funding constraints for continued development.
Tanzania's Electronic Tax Stamps (ETS) system — used to track excisable products such as tobacco, beer, spirits, and non-alcoholic beverages — is vulnerable to fraud and evasion if stamps are forged or products circulate outside the official tax net. The Tanzania Revenue Authority, working with a private sector vendor (SICPA), deployed an AI-powered market surveillance initiative to enhance the ETS system through real-time monitoring and fraud detection. The system captures data from multiple sources including production records, scan location data, cameras, and sensors, then processes this using AI algorithms and expert insights for both passive and active fraud detection. Automatic fraud alerts are generated and legal insights recorded for TRA enforcement teams. By end of 2024, over 10.38 billion products had been marked under the system. The initiative has enabled enforcement teams to flag and trace marked products both within Tanzania and across borders, improving enforcement strategy and operational efficiency. The main challenge cited is technical skills — maintaining and developing AI capabilities within the revenue authority. The initiative is described as replicable in any country using secure product traceability systems for domestic revenue mobilization on excisable goods.
The Korea Customs Service handles a high and growing volume of public inquiries about customs procedures, regulations, and individual case status — driven in large part by the rapid growth of overseas direct purchasing among Korean consumers. Responding to these inquiries manually is time-intensive and creates delays for users. An AI-powered automated customer service system was developed using historical public complaint and inquiry data to train a model capable of providing responses comparable in quality to those of human expert agents. The system processes incoming inquiries through an automated response pipeline, reducing staff workload while maintaining service quality. Primary data inputs are historical public inquiry records and complaints, alongside customs regulations and procedural documentation. Reported outcomes include a measurable reduction in staff workload while sustaining high-quality responses for citizens and businesses navigating customs requirements — particularly relevant as the volume of individual import transactions continues to grow.
South Korea's National Tax Service receives a high volume of public inquiries about comprehensive income tax — a domain characterized by complex rules, numerous exceptions, and interactions between multiple legal provisions that standard keyword-based or rule-based chatbots have historically struggled to handle accurately. The NTS replaced its previous rule-based chatbot with a system built on HyperClova X (a Korean hyper-scale language model) capable of handling complex, multi-turn conversations about income tax. The system is trained directly on original legal statutes and uses Graph DB-based reference relationship processing to systematically map connections between related legal provisions, enabling it to reason through complex tax scenarios with greater accuracy than conventional retrieval approaches. A key design priority was reducing AI "hallucination" — the tendency to generate plausible but incorrect information — by grounding all responses firmly in statutory text. Primary data inputs are income tax legislation, legal statutes, and regulatory guidance. The system reduces response failure rates compared to its predecessor and provides citizens with reliable, legally grounded answers to complex tax questions, reducing escalations to human agents.
Citizens interacting with Albanian public services have historically faced bureaucratic complexity, long wait times, and limited access outside of office hours. The Virtual Assistant "Diella," integrated into the e-Albania platform, uses generative AI to provide 24/7 real-time assistance through voice or text, allowing citizens to navigate public services, ask questions, and obtain official documents — complete with electronic seals — through natural conversational interaction. The system automates responses to citizen inquiries and document generation, reducing administrative burden on public institutions. Technically, Diella combines speech-to-text, text-to-speech, NLP for intent understanding, and avatar movement coordination in a conversational AI system. All development required significant customization for the Albanian language — a technically demanding task given Albanian is a low-resource language with limited existing AI training data. Since launch, Diella has responded to over one million citizen inquiries, with over 300,000 responses delivered in the first months of 2025 alone. It supports the generation of over 54 types of electronically sealed official documents. Diella 3.0 — the current version — enables full interaction via voice commands, including completing service requests through conversational speech. The biggest challenge was adapting and integrating four advanced AI modules (text-to-speech, speech-to-text, avatar coordination, command execution) into the Albanian language — a linguistically complex undertaking for a less-resourced language.
Blood banks in Brazil and globally face persistent shortages driven by irregular donation patterns and ineffective outreach to first-time and infrequent donors, who are the hardest to convert into regular contributors. This AI-powered initiative — originally designed and implemented at Fundação Hemominas by a public health servant and developed with Xertica.ai using Google Cloud and Gemini AI — deploys an AI agent that proactively contacts potential blood donors through WhatsApp, phone calls, and other digital channels. The system monitors real-time blood stock levels across hemocenters to identify shortages and dynamically sets data-driven outreach goals, ensuring contact happens precisely when and where donations are most needed. It personalizes communication based on donor profiles, donation history, and past engagement patterns, aiming to reach each person through channels they already use in a way that feels human and respectful rather than automated. The primary data inputs are blood stock inventory data and donor engagement data. Since the pilot began, nearly 15% of contacted individuals proceed to schedule a donation — a strong conversion rate for public health outreach. The biggest challenge was building trust with first-time and infrequent donors skeptical of automated messaging, requiring careful communication design and continuous adaptation. The initiative's creator advocates for this to become a national program across Brazil's SUS health system.
Public safety in Ethiopia — as in many countries — is constrained by the difficulty citizens face in reporting incidents to the police in a timely, convenient, and trustworthy manner. The Ethiopian Federal Police Citizen Engagement App (EFPApp), developed in-house by EAII, provides a mobile platform through which citizens can report incidents, crimes, and suspicious activities directly to the Ethiopian Federal Police. Key features include multimedia submission (photos, videos), geolocation tagging for precise location reporting, and real-time status updates on reported cases. On the police side, AI aggregates and analyzes incoming reports to detect emerging crime patterns, prioritize responses, and optimize resource allocation. The system relies primarily on citizen-generated incident data, including multimedia attachments and geospatial information. The tool is fully operational and has improved police response times, enhanced communication between citizens and law enforcement, and supported AI-driven crime pattern detection. On May 16, 2025, the EFPApp won first place in the Best Police App category at the International Police Summit in Dubai — international recognition of its innovation and effectiveness. The single biggest challenge is sustainable funding: while initial development was supported, ongoing costs for maintenance, upgrades, user training, and expansion require consistent financial resources that are difficult to secure.
Municipal governments receive thousands of citizen applications, complaints, and requests across digital channels every day. Manually reading, interpreting, and routing each submission to the correct department is labor-intensive, slow, and prone to human error — leading to misrouted applications, service delays, and citizen dissatisfaction. Bağcılar Municipality, building on a secure data infrastructure established under the World Bank-supported ISMEP project, developed an AI-powered application classification and routing engine integrated into its Electronic Management Information System (EYBS). Using NLP, the system semantically analyzes the free-text content of each incoming citizen submission, classifies it according to the relevant municipal service, and automatically routes it to the correct directorate within seconds — without human intervention. The AI model continuously learns from each classification and any corrections made by staff. Training data consisted of tens of thousands of labeled historical application records. Since implementation, routing accuracy has exceeded 95%, the process that previously took minutes to hours now takes seconds, workforce previously dedicated to manual routing has been redeployed to higher-value tasks, and real-time reporting provides management with data for resource planning. The project received IDC and TBD (Informatics Association of Turkey) awards in 2022 and 2023. The biggest challenge was not the technology itself but the data engineering: processing the "chaotic" free-text of citizen applications — full of slang, dialects, spelling errors, and colloquialisms — required intensive data labeling and model training, and complete mapping of hundreds of service types to responsible directorates.
Citizens seeking information from Bağcılar Municipality have historically faced the limitations of traditional service channels: restricted office hours, call center queues, and the need to navigate complex websites or visit multiple departments. BAGBİ is an AI-powered chatbot that serves as a 24/7 autonomous digital assistant, offering citizens personalized, real-time municipal information through a conversational interface. Connected to the municipality's centralized Management Information System (built on World Bank ISMEP-supported infrastructure), BAGBİ functions as a "single source of truth" — citizens can verify personal debt status, query zoning information and property valuations, check garbage collection schedules, find on-duty pharmacy locations, and explore cultural events through natural language conversation. The system uses NLP with a convergence methodology to handle typographical errors and colloquial phrasing, and continuously improves through machine learning from user interactions and feedback. Data inputs include citizen records from institutional systems and a rich FAQ dataset (430 FAQs, 1,900 alternative questions, 2,800 keywords, and 8,500 pieces of user feedback). The message understanding rate stands at 83%, and usage grew from 320,000 messages in 2023 to 783,000 in 2024 — a 144% increase — across 250,000 unique users. Workforce savings equivalent to approximately 570 workdays have been achieved. The greatest challenge was institutionalizing the knowledge: teaching the AI the full range of the municipality's services required active participation from 25 different directorates, significant data structuring, and sustained organizational coordination.
Navigating the full catalog of government services on a ministry portal can be confusing and time-consuming for citizens and businesses who may not know the precise terminology for what they are looking for. The Ministry of Economy and Tourism addressed this through a ChatGPT-powered AI search function, positioned alongside the regular search bar on the ministry's portal. Citizens can describe what they want to do in natural language, and the AI chatbot interprets their request and redirects them to the correct service destination — effectively serving as an intelligent navigation layer over the existing service catalog. The system uses NLP and generative AI (GPT-based) for text analysis and response generation, drawing on citizen records and service data. The tool is fully operational and has improved both customer satisfaction and process efficiency. The primary challenge has been infrastructure — specifically the absence of an on-premises solution and reliance on cloud infrastructure, which raises questions around data sovereignty and continuity of service.
The 2022 Itaewon crowd crush, in which 159 people died during a Halloween festival, exposed a critical gap in South Korea's public safety infrastructure: the absence of real-time crowd density monitoring capable of triggering early warnings to authorities and citizens. In response, the government developed the Crowd Management Support System, which uses anonymized mobile carrier data to monitor crowd density in real time across designated high-risk locations during major public events. When crowd density exceeds pre-defined risk thresholds, the system automatically sends alerts to local governments and police to enable coordinated crowd management responses, and can also dispatch direct mobile alerts to users in affected areas. The system has been deployed at events including the Busan Fireworks Festival. Primary data inputs are real-time anonymized mobile carrier location data providing density estimates. The initiative directly connects government agencies, law enforcement, and citizens through a shared real-time risk awareness system — a model of AI-enabled proactive disaster prevention that emerged from a documented public safety failure.
Effective environmental governance at the watershed level requires ongoing engagement between public authorities and local communities, yet traditional consultation mechanisms are infrequent, one-directional, and inaccessible to many citizens. An Integrated Environmental Information Platform was developed for the Miho River basin to serve as a digital hub for environmental monitoring and structured citizen participation. A central feature is an Environmental Complaint Service through which citizens can submit complaints about environmental issues in their area, share opinions, and engage with posts from other community members — fostering a model of public-private community governance around environmental management. AI underpins the platform's ability to aggregate, organize, and surface citizen inputs alongside environmental monitoring data. Primary data inputs are citizen-submitted environmental complaints and engagement records, alongside environmental monitoring data for the basin. The initiative demonstrates how AI-enabled digital platforms can institutionalize continuous citizen participation in environmental governance, moving beyond periodic formal consultations to an always-on, community-driven feedback and accountability channel.
Traditional tourism data collection in Argentina — primarily through surveys such as the Encuesta de Viajes y Turismo de los Hogares (EVyTH) — is slow, costly, and unable to capture granular or real-time information on domestic travel patterns. This initiative integrates anonymized mobile device data (Identifier for Advertising — IFA) with traditional survey data to provide significantly more detailed and timely insights into tourist flows, movement patterns, origin-destination pairs, and duration of visits. Co-developed with the private sector and academia, the pilot demonstrated the potential for mobile data to fill critical gaps — particularly for spontaneous, short-term, or non-resident travel that surveys typically miss — and to support seasonal planning, destination management, and evidence-based tourism policy. While the proof of concept was successful, the initiative did not progress to full implementation due to challenges around data privacy and regulation, limited institutional capacity to analyze and interpret large-scale non-traditional data, and the absence of a long-term sustainability framework. The biggest challenge was balancing innovative data use with regulatory, technical, and institutional constraints — all of which limited the project's continuity.
Public service digitization in Brazil has traditionally required significant technical expertise and long implementation timelines, placing it out of reach for municipalities with limited IT capacity. Colab AI's Public Services Creator addresses this by providing a generative AI platform that allows public managers to create new digital services — complete with citizen-facing journeys and internal government workflows — simply by typing a natural language prompt. For example, entering "create a preschool enrollment service prioritizing vulnerable families" generates a ready-to-deploy service with forms, validations, notifications, and approval workflows across WhatsApp, mobile apps, and portals. The system draws on hundreds of real public sector use cases to generate contextually appropriate outputs, using NLP and intelligent process automation. In Niterói, the platform enabled critical services such as early childhood enrollment to be launched in minutes rather than months, expanded access for underserved families via WhatsApp, and reduced the dependence on technical teams for workflow design. The technology was provided primarily by a private sector vendor (Colab). The biggest challenge was institutional resistance and low initial familiarity with generative AI among public managers, overcome through piloting, training, and demonstrating early wins. The platform is already used in multiple Brazilian cities and is described as scalable to low-resource environments globally.
Governments often struggle to discover and procure innovative technology solutions from startups, while startups lack visibility into government procurement opportunities. GovTech Place is Brazil's first B2G (Business-to-Government) digital marketplace, designed to bridge this gap by mapping and organizing innovative solutions available in the market, simplifying public procurement processes, and helping public managers identify the technologies that best meet their needs. AI is used for matching, classification, and automating decision-making within the platform. Primary data inputs are innovative public procurement data from governments and startups. Funded by a British Government grant and developed through a public-private collaboration, the platform has been formally launched and is establishing itself as a hub for Brazil's GovTech ecosystem. Key impacts include reduced complexity in procurement processes, increased visibility for startups, and support for public sector modernization. The greatest challenge has been overcoming legal and cultural barriers to contracting startups in the public sector, compounded by issues of data standardization, technical training needs for public managers, and institutional resistance to change. The platform's modular structure is designed for global scaling and adaptation to different regulatory environments.
AI-PACT is a European Digital Innovation Hub (EDIH) focused on facilitating the adoption of AI in public administrations, SMEs, and startups across Italy. As part of the EU's EDIH network — a key instrument for supporting the digital transformation of public sector entities — AI-PACT's core initial activity is mapping the Italian GovTech and AI ecosystem, specifically cataloging startups and solutions relevant to public sector AI adoption. Primary data inputs at this stage are startup data and ecosystem information. The initiative is a collaboration between SDA Bocconi and private sector and academic partners. As of submission, the mapping activity had not yet formally begun, and the key challenge identified was determining the correct characteristics and criteria for mapping startups effectively. The initiative represents a supply-side approach to AI in government — building the knowledge infrastructure and network that enables public administrations to identify and adopt AI solutions — rather than a direct AI deployment in a public service.
Argentina faces a significant challenge in primary education: approximately 46% of 3rd graders do not understand what they read. This reading comprehension gap, if unaddressed, compounds over time and constrains educational attainment and economic opportunity. DidactIA is a generative AI-powered educational tool that acts as a personalized, interactive reading tutor for primary school students. Rather than simply evaluating answers, the AI guides students through the process of constructing correct responses, adapting reading activities to each student's pace and level through curriculum-aligned, adaptive learning experiences. For teachers, the platform allows easy configuration of interactive literature-based activities and generates detailed performance reports to inform instruction. Data inputs include literary texts, reading materials, and student interaction data including performance logs and activity tracking. Co-developed with private sector and academic partners, pilot testing has demonstrated improved student engagement, stronger differentiated instruction, and the tool's ability to help teachers identify and address learning gaps. Academic partners have validated the pedagogical approach. The most significant challenge is infrastructure: many Argentine schools lack sufficient individual devices or stable internet connectivity to fully leverage the tool's personalized, student-centered approach — a constraint that directly affects scalability and equitable implementation across diverse educational settings.
Secondary school dropout in Argentina — as in much of Latin America — represents one of the most persistent and costly failures of the education system. The Early Warning System (SAT) in Mendoza uses machine learning to predict, at the individual student level, the probability of dropping out before the next school year. Drawing on student administrative records (enrollment, attendance, grades) and socio-demographic data (parental education level, family background) integrated from the provincial GEM education management system, the model classifies students into risk tiers (low, medium, high), identifies likely causes of dropout risk for each student, and suggests targeted interventions. School leaders receive actionable insights and can track intervention responses over time. Co-developed with Fundar and academic partners, the system makes Mendoza the first Argentine province to implement a predictive early warning system in education. Reported outcomes include earlier identification of at-risk students, reduced administrative burden on schools through automated risk detection, improved coordination between schools, provincial authorities, and technical experts, and a precedent for evidence-based educational management. The biggest challenge was ensuring data interoperability and confidentiality across multiple institutions and systems — particularly given the sensitivity of data involving minors — which required strong institutional coordination, legal safeguards, and robust technical infrastructure.
Reviewing architectural building plans for compliance with health facility regulations is a time-intensive, expert-dependent process that creates bottlenecks in the approval of new healthcare infrastructure. The Ministry of Health and Prevention is developing an AI-powered automated plan review system that uses computer vision to analyze architectural blueprints and NLP to assess compliance with relevant regulatory frameworks. The system scans submitted plans against a compliance framework and health facility standards, flagging areas of non-compliance for human review. Primary data inputs are the compliance regulatory framework and architectural building plans in digital format. The initiative was co-developed with the private sector and is currently in pre-production. Expected impact is a significant reduction in service time for architectural plan approvals. The primary challenge is data quality — ensuring that the AI models are trained on sufficient, accurate, and representative examples of compliant and non-compliant building plans to achieve reliable automated assessment.
Early and accurate detection of critical conditions such as breast cancer, tuberculosis, and stroke is essential for patient outcomes, yet radiology workflows in busy healthcare systems can be slow, inconsistent, and prone to human fatigue. Emirates Health Services (EHS) has integrated AI-powered diagnostic imaging tools into its radiology departments to address these challenges. The AI tools automatically analyze medical images — mammograms, chest X-rays, and brain CT scans — in DICOM format to detect signs of disease, highlight abnormal findings, and prioritize urgent cases for radiologists. Integrated into existing RIS/PACS systems, the tools streamline reporting workflows and reduce diagnostic turnaround times. Technology was provided primarily by private sector vendors. The outcomes reported are among the most quantitatively striking in this collection: breast cancer detection accuracy improved from 85% to 98%; tuberculosis detection improved from 80% to 99%; false negative rates for tuberculosis fell from 0.06% to 0.015%; recall rates for breast cancer screening dropped from 10.09% to 4%; and diagnostic turnaround for breast cancer fell from 14 days to 5 days, with tuberculosis diagnosis reduced from 72 hours to 10 hours. The initiative received the "Future Fit" sign from the UAE Government Development and Future Office in 2024. The biggest challenge was institutional resistance and clinical adoption: radiologists were initially hesitant to trust AI outputs, and concerns over medico-legal accountability required phased implementation, stakeholder engagement, and a clear positioning of AI as a clinical support tool rather than a replacement for professional judgment.
Italy's PagoPA serves as a critical digital infrastructure provider, managing certified digital communications (SEND), developer-facing documentation portals, and the national AppIO interface through which citizens interact with government services. Three distinct AI initiatives are bundled under this submission. First, the Smart Verification tool for SEND uses OCR and machine learning to automatically extract key data from postal receipt images (sender, date, signature check) and cross-validates this against system metadata, flagging discrepancies for review — significantly reducing manual workload and improving the consistency of document validation in the certified digital delivery system. Second, Discovery is a generative AI chatbot integrated into PagoPA's Developer Portal that understands technical queries from developers and guides them through official documentation to resolve problems — reducing support burden and improving technical adoption. Third, the Smart Agent for AppIO is a non-generative AI agent that helps citizens resolve issues related to accessing AppIO using their digital identity (SPID/CIE). Data inputs include citizen records and official government texts. Collectively these tools have achieved a significant reduction in false positive and negative rates (for the SEND verification), though precise outcome metrics are still being calibrated. The challenges varied by component: for SEND, data quality in the underlying receipt texts; for Discovery, the quality and currency of source documentation; and for the AppIO agent, convincing stakeholders to invest in an AI solution predating the generative AI wave.