AI in Public Administration: Global Experience

Brief summary: Artificial intelligence in public administration is moving from the pilot stage to systemic integration. This article provides an analysis of global regulatory models, real-world implementation cases from Singapore, Estonia, the UAE, and Kazakhstan, a functional review of how AI is used across public-sector functions, as well as an honest discussion of risks and ethical challenges.

Contents

1. From e-government to smart government
2. Three regulatory models: the EU, the US, and the "third way"
3. Singapore: how a small country became a global benchmark
4. Estonia: the invisible state and agentic AI
5. UAE: when the goal is an AI-native government by 2027
6. Kazakhstan: a proactive state and the "Digital Family Map"
7. Where AI is already working in the public sector: a functional review
8. Systemic risks and ethical challenges
9. What this means for the Russian market
10. FAQ

1. From e-government to smart government

In the mid-2020s, the concept of the digital state is undergoing a fundamental transformation. If the first wave of digitalization—the conventional period from 2000 to 2018—was reduced to moving paper-based regulations into online forms, today we are talking about a qualitatively different paradigm.

Artificial intelligence in public administration has ceased to be a technological experiment and has become an architectural solution. According to the OECD, about 67% of member states are already using AI technologies in the design or delivery of public services. At the same time, most initiatives are still at the pilot stage—full systemic integration remains a task for the next five years.

What exactly is changing

The OECD identifies six dimensions of digital government maturity: digital by design, data-driven, government as a platform, open by default, user-driven, and proactive. AI is a catalyst for all six—but in different ways.

It is important to note the key shift: the state is ceasing to be reactive—responding to citizens' requests—and is becoming proactive, independently identifying needs and preventing problems before they arise.

2. Three regulatory models: the EU, the US, and the "third way"

Before examining the cases, it is important to understand the context: countries approach AI regulation in the public sector in fundamentally different ways. These are not just legal differences—they are different philosophies of the relationship between the state, technology, and the citizen.

European Union: a risk-based imperative

The EU was the first in the world to adopt a comprehensive AI law (the EU AI Act) with extraterritorial effect. The logic is simple and strict: classify systems by risk level.

Completely prohibited — social scoring systems, mass biometric surveillance in public places without exceptions. Rationale: the potential for discrimination is incompatible with the concept of human dignity.

High risk (strict oversight) — AI in law enforcement, immigration control, education, and the distribution of social benefits. Mandatory audits, human oversight, and registration in a centralized EU database are required.

The European model is slower to implement, but it creates a long-term foundation of trust. GDPR, in conjunction with the AI Act, effectively forces developers to implement the principle of "ethics by design" at the system architecture level.

US: decentralization and speed

The American approach is the opposite of the European one. A fragmented sectoral model: each federal agency develops its own policy, and states experiment independently.

Executive Order 14110 instructed more than 50 agencies to develop individual AI policies. At the same time, a deregulatory vector is accelerating the diffusion of technologies throughout the economy. Examples of targeted regulation: Colorado's AI law (core public services), New York's law on auditing bias in automated hiring tools.

The American model accepts higher regulatory risks in exchange for speed. This is a conscious choice: technological leadership is more important than standardization.

The third way: pragmatic sovereignty

Singapore, Estonia, the UAE, and Kazakhstan are forming a third model—not a copy of either Brussels or Washington. Its features are strong state will, a focus on sovereign data infrastructure, a pragmatic balance between efficiency and rights protection, and an iterative approach without rigid prior regulation.

3. Singapore: how a small country became a global benchmark

Singapore ranks second in the global digital competitiveness ranking in 2025. With a population of fewer than 6 million people and the physical constraints of a small island, the country has turned digital transformation into a national survival project.

The national AI strategy is integrated into the Smart Nation concept—a joint creation of innovation by government, business, and citizens. This is not just a nice slogan: it is about real co-creation at the level of architectural solutions.

Healthcare: AimSG as an App Store for doctors

The AimSG platform works on the principle of an app store for medical professionals. A doctor selects a verified AI algorithm for a specific task—X-ray analysis, predictive diagnosis, triage. All algorithms are validated by the regulator, and all decisions are documented.

Key result: the primary chest X-ray analysis model has been deployed in all public hospitals and provides automatic triage of critical cases. The specialist receives priority flags before even viewing the image—and immediately focuses on the most dangerous pathologies.

Transport: autonomous transport as urban infrastructure

Since 2015, the Land Transport Authority (LTA) has been systematically developing driverless transport. Today it is not an experiment—it is part of urban infrastructure. Driverless buses on dedicated routes, Robosweeper cleaning robots in the Esplanade tourist areas. Every autonomous vehicle undergoes certification at the CETRAN test center, which simulates real complex road conditions.

Border control: passport-less immigration

Perhaps the most illustrative case from the UX perspective. The facial and iris recognition system reduces border control processing time by 40%. By the end of 2024, the technology had been deployed across all Changi Airport terminals. A resident passes through without physical documents—the system identifies them automatically via biometrics.

AI Verify: trust infrastructure as an export product

Special attention should be paid to the AI Verify toolkit — the world's first open AI testing system for compliance with 11 principles: transparency, explainability, robustness against attacks, and others. This is not just an internal regulatory tool — it is an export product that Singapore promotes as an international standard.

Singapore's key lesson: Public trust in AI is built not through declarations, but through verifiable technical standards. AI Verify allows a developer to document and prove the system's safety before deployment.

4. Estonia: the invisible state and agentic AI

Estonia is perhaps the most interesting case for a technical audience. A country with a population of 1.3 million has built the world's most functional digital state ecosystem — based on the X-Road platform and end-to-end digital identification.

In 2024–2025, Estonia took the next step: moving from chatbots to a network of interoperable virtual agents called Bürokratt.

Bürokratt: from Rasa to LLM

Bürokratt is not a single product. It is an interoperable network of public and private AI solutions that, from the user's perspective, functions as a single access channel to services.

The technical evolution is telling. The first versions were based on the Rasa platform and required labor-intensive manual training on predefined question–answer pairs. Any change in regulations meant manual rewriting of dialogs. In 2024, a mass transition to LLMs began: the system understands semantic context without prior preparation and retrieves up-to-date information from agency databases in real time.

The 2026 vision: every agency has its own personalized AI agent. The agents interact with each other autonomously — for example, a benefits application and a municipal appointment booking can be submitted at the same time through a single request.

Automating justice: three tools

Estonia has introduced AI into the judicial system to address the problem of overloaded courts.

Salme — a speech recognition system specifically trained on Estonian legal language (800 hours of audio recordings, 800 million words). Automatic real-time transcription of hearings.

Krat — automatic anonymization of court decisions before publication. The algorithm identifies and removes participants' personal data. Result: the principle of public access to justice is upheld without violating privacy.

Semi-automatic orders — for low-value claims involving debts and alimony, the algorithm checks jurisdiction and formal criteria, after which it generates a payment order with the force of an enforcement writ.

At the same time, Estonia's expert community and the Supreme Court consistently defend the principle that the “human judge” remains the final link in any significant case. AI is a tool for relieving the load, not a replacement.

Estonia's key lesson: Architectural interoperability from the very beginning (X-Road) is a strategic advantage that makes it possible to add an AI layer without rewriting the underlying infrastructure.

5. UAE: when the goal is an AI-native government by 2027

The UAE is the most ambitious example in terms of declared goals. The AI Strategy for 2031 forecasts that by 2030 AI will bring the region about $320 billion, generating up to 14% of the country's GDP through intelligent technologies.

Each federal ministry in the UAE has a Chief AI Officer position — a dedicated role responsible for implementing the AI strategy and developing employees' competencies.

Dubai Customs: Dubai Customs 2030 strategy

Risk Engine — a risk analysis engine based on ML algorithms. The system assesses customs declarations in real time, correlating data from global and local sources. The goal: automatic detection of smuggling, IP violations, and goods-origin fraud.

TruRisk — an AI tool for AML compliance. Implementation results: a 44% reduction in false positives in anti-money-laundering checks and a 70–80% reduction in manual compliance labor. These are not marketing figures — they are key operational KPIs for the customs service.

Smart police stations: 27 locations without staff

Dubai has deployed a network of 27 fully automated police stations operating 24/7 without on-site personnel. Citizens receive 45 services in seven languages — from filing crime reports to processing documents. AI verifies identity and routes requests.

The Oyoon surveillance system combines more than 5,000 cameras with face recognition and behavior analysis capabilities. Predictive policing uses historical data to build crime “heat maps” — the algorithm identifies zones with a high likelihood of offenses and optimizes patrols.

The UAE Public Prosecution announced plans to use VR to recreate crime scenes — this allows courts to analyze the sequence of events in detail without physical presence at the site.

The UAE's key lesson: Political will and centralized governance accelerate adoption, but they also create heightened demands for ethical auditing and mechanisms for citizens to challenge decisions.

6. Kazakhstan: a proactive state and the “Digital Family Map”

By 2025, Kazakhstan had risen to 24th place in the UN Global E-Government Index — an impressive result for a country that was not in the top 50 ten years ago. The main direction of the national strategy is the shift from an application-based service model to a proactive one.

Digital Family Map: AI instead of queues

The “Digital Family Map” built on the Smart Data Ukimet analytics platform is probably the most significant social AI project in the post-Soviet space.

Mechanics: The system automatically analyzes more than 100 socio-economic indicators for each family in the country. The citizen does not need to contact a government agency — the state itself identifies eligibility for a benefit or assistance. The notification arrives by SMS, and consent to receive the payment is given by reply SMS.

Results by mid-2025:

• 4.5 million proactive services delivered
• 52,000 people significantly improved their well-being
• 90,000 improper entries in housing waiting lists identified and removed
• The efficiency of jury selection increased sixfold

Sergek: when safety is measured in percentages

The Sergek public safety and traffic management system combines a network of cameras and sensors in Astana and Almaty. Real operational results: a 40–50% reduction in road traffic fatalities and a 40% reduction in crime in public places.

Technically: face recognition, vehicle trajectory tracking, and automatic 24/7 detection of traffic violations. By the end of 2025, a module for adaptive traffic light control is planned to be launched, with a target of reducing congestion by 15–20%.

In July 2025, Kazakhstan launched the most powerful supercomputer in the region based on NVIDIA H200 — sovereign computing infrastructure for national AI projects.

Kazakhstan's key lesson: A proactive service model requires high-quality government data. Investments in dataset quality are investments in the accuracy and fairness of AI solutions.

7. Where AI is already working in the public sector: a functional analysis

An OECD study covering more than 200 cases across 11 functional areas makes it possible to systematize the real-world use of AI in government.

Tax and financial administration

Tax authorities are moving from sample audits to comprehensive transaction monitoring. ML algorithms identify tax evasion patterns invisible to traditional methods. AI in public financial management improves the accuracy of budget revenue forecasting and minimizes the human factor in the allocation of limits.

Public procurement and anti-corruption

Procurement is one of the most corruption-prone areas of government. AI is transforming the entire procurement cycle:

• Preventive monitoring: analysis of relationships between tender participants, identification of signs of cartel collusion and supplier affiliations with officials.
• Price analysis: comparison of market quotes with bid prices, preventing purchases at inflated costs.
• Execution management: tracking contract deadlines and quality, and generating supplier trust ratings.

Policy evaluation and civic participation

NLP models analyze social networks, open data, and survey results in real time. The government gains the ability to adjust policy initiatives already at the implementation stage. AI moderation of civic participation platforms helps classify thousands of public proposals and identify constructive ideas.

Healthcare and social protection

In addition to the cases already discussed, these include predictive identification of risk groups (for example, citizens with a high probability of diabetes or cardiovascular disease), optimization of patient routing in overloaded systems, and automatic verification of the legitimacy of social benefits.

8. Systemic risks and ethical challenges

It would be dishonest to talk only about successes. The integration of AI into public administration entails systemic risks that everyone involved in such projects must be aware of.

Algorithmic bias: the problem of the "black box"

AI models are trained on historical data—and historical data reflect historically established inequality. Systems in law enforcement, social benefits, or housing allocation risk reproducing discriminatory patterns at industrial scale.

The problem is amplified by the "black box": in deep-learning-based models, it is impossible to explain why the algorithm made a particular decision. In the context of assigning social benefits or issuing a court ruling, this creates not just a technical issue, but a constitutional one.

The legal response is the concept of the "right to explanation," implemented in GDPR and the EU AI Act. The technical response is the development of XAI (Explainable AI) as a mandatory component of government AI systems.

The digital divide as a new form of social exclusion

According to the OECD, the gap in the use of generative AI between age groups in developed countries exceeds 53 percentage points. If government services become available primarily through intelligent interfaces, part of the population—older citizens, people with disabilities, and residents with low digital literacy—risks being left outside the social security system.

Inclusive design of government AI services is not an optional feature, but a condition of legitimacy.

The paradox of inaction

A less obvious, but recognized, risk: refusing to implement AI is itself a risky decision. The degradation of public services relative to the private sector leads to a decline in public trust in institutions, a "drain" of interactions into private ecosystems, and a loss of strategic data. Governments that delay integration risk becoming irrelevant to the technically advanced part of the population.

Data sovereignty risk

The use of foreign LLM platforms and cloud services in critical government functions creates dependence on foreign infrastructure. That is why Kazakhstan is investing in a sovereign supercomputer, Russia in the development of GigaChat and YandexGPT, and the EU in the GAIA-X initiative.

9. What this means for the Russian market

The Russian context creates specific conditions for implementing AI in public administration.

At the regulatory level, the national project "Data Economy" is in effect, setting the direction for digital transformation through 2030. Regulatory experiments (sandboxes) for AI provide a limited opportunity to test technologies outside standard legal frameworks.

At the infrastructure level, domestic LLM platforms play a key role—YandexGPT, GigaChat, as well as a number of specialized solutions. Restrictions on access to Western GPUs create demand for optimization for available hardware (Huawei Ascend, domestic accelerators in the future).

From the standpoint of applied tasks, the Russian public sector shows strong interest in: automating work with citizen inquiries, intelligent monitoring of public procurement and contract execution, NLP processing of regulatory documents, and predictive analytics in the social sphere—following a model similar to Kazakhstan's "Digital Family Map."

For teams working at the intersection of AI and the public sector, the following are critically important: understanding regulatory constraints, experience integrating with SMEV and government APIs, and the ability to explain how algorithms work in language that is understandable to regulators.

10. FAQ

What is a "Smart State" and how does it differ from e-government?

E-government is the transfer of existing processes into a digital format (online applications, electronic documents). A Smart State is a qualitatively different level: the use of AI, big data, and predictive analytics to transform the very logic of public administration. The key difference is the shift from a reactive to a proactive model, in which the state independently identifies citizens' needs and offers solutions without being contacted first.

Which countries are leading in the implementation of AI in public administration in 2025?

By the combined indicators of systematic approach, scale, and effectiveness of implementation, the leaders are Singapore (2nd place in the IMD 2025 digital competitiveness ranking), Estonia (a pioneer of agentic AI in the public sector), the UAE (the most ambitious transformation program), and Kazakhstan (a regional leader in Eurasia, 24th place in the UN ranking).

What are the main risks of implementing AI in public administration?

The main risks are: algorithmic bias (the reproduction of historical discrimination in systems that make decisions about citizens), the opacity of the "black box" (the inability to explain a specific decision), digital exclusion (the inaccessibility of intelligent interfaces for part of the population), and the risk to data sovereignty (dependence of critical infrastructure on foreign platforms).

How does the proactive model of government services based on AI work?

Instead of waiting for a citizen to submit an application, the system automatically analyzes their profile across many parameters (income, family composition, employment, health, etc.), identifies eligibility for a benefit or assistance, and initiates an offer. Kazakhstan’s “Digital Family Card” is the most comprehensive implemented example: the system analyzes more than 100 indicators for each family and notifies the citizen via SMS.

What is Explainable AI (XAI), and why is it important for the public sector?

XAI is a field of AI development whose goal is to make algorithmic decisions interpretable for humans. In the public sector, this is critically important: a citizen who has been denied a benefit or placed in a queue has the right to know the reason. The EU AI Act legally enshrines this requirement for high-risk systems. Without XAI, introducing AI into legally significant processes is a source of systemic legal risks.

Can AI be introduced into public administration without special legislation?

Technically, yes — within regulatory sandboxes or pilot projects in functions that do not directly affect citizens’ rights. However, sustainable large-scale implementation requires a legal framework: to protect citizens from incorrect algorithmic decisions, to define liability for system errors, and to regulate the use of personal data.

Conclusion

Artificial intelligence in public administration is not a technological trend, but a structural transformation that is happening regardless of whether you are involved in it. Leading countries have already passed the stage of “romanticism” and moved on to pragmatic systemic integration.

Key takeaways from this analysis:

• Trust is built with technical tools, not declarations — the example of Singapore with AI Verify.
• Architectural decisions made today determine tomorrow’s capabilities — X-Road interoperability gave Estonia a 20-year advantage.
• Proactivity requires high-quality data — investments in datasets are more important than investments in algorithms.
• Computational sovereignty is becoming a component of national security — alongside territorial sovereignty.

The next stage is agentic AI, capable not only of advising but also of autonomously performing complex administrative tasks. The role of a public sector specialist is being transformed: from a rules executor to an architect and auditor of intelligent systems.

This material was prepared based on analytical data from the OECD, reports from the IMD Digital Competitiveness Ranking 2025, and open sources from the governments of Singapore, Estonia, the UAE, and Kazakhstan.