Public administrations face the same equation as companies, with less room for manoeuvre: rising service expectations, budgets that do not follow, retirements that are hard to replace. AI answers part of it, and the movement is accelerating: the US federal inventory tripled in one year to 2,133 use cases, and a full-scale trial in the United Kingdom measured 26 minutes saved per civil servant per day. Here is what AI already does in public services, what it returns, and the framework to respect, seen from Belgium.
The public sector is no longer a spectator. In the United States, the consolidated 2024 federal inventory lists 2,133 public AI use cases across 41 agencies, up from 710 a year earlier: a threefold increase in one year. The Department of Health and Human Services alone reports 271, and 227 uses are classified as rights- or safety-impacting, hence subject to reinforced safeguards.
In the United Kingdom, the Government Digital Service ran from 30 September to 31 December 2024 one of the largest controlled generative AI trials in the world: 20,000 civil servants across 12 organisations used Microsoft 365 Copilot to draft, summarise and search. Measured average gain: 26 minutes per day per official, nearly two weeks a year, and 82% of participants did not want to go back.
The OECD, finally, published on 18 September 2025 its report Governing with Artificial Intelligence: of 200 use cases documented across 11 functions of government, 57% serve to automate, streamline or tailor services, 45% support decision making and forecasting, 30% strengthen accountability and anomaly detection. The report also notes fragmented adoption and still marginal use of large language models: the room for progress remains considerable.
Belgium moves step by step. The national convergence plan for the development of AI, approved on 28 October 2022, sets nine objectives and aims to make the country a « Smart AI Nation », with the AI4Belgium coalition as rallying point. On 11 July 2025, 40 federal organisations signed the charter for responsible use of AI in public services, coordinated by the FPS BOSA: transparency, human oversight and respect for the citizen are set as principles before scaling up.
Behind the numbers, the uses are concrete and often less spectacular than imagined. The cases that work share one trait: a high volume of repetitive tasks, where every minute saved multiplies across thousands of files.
The UK trial provides the most solid measurement to date: 26 minutes per day, concentrated on searching for information (26 minutes), writing emails (25 minutes) and summarising documents (24 minutes). These are individual productivity gains, not yet budget savings: they convert into value if the freed-up time is reinvested in supporting users or clearing backlogs.
The OECD tempers the enthusiasm: adoption remains fragmented and uneven, concentrated on high-transaction services; projects run into skills, legacy systems, data sharing and budgets. The report identifies seven enablers to move from pilot to scale: governance, data, digital infrastructure, skills, investment, procurement and partnerships. In other words, technology is rarely the limiting factor; the organisation is.
Start with low-risk, high-volume internal uses (summarising, drafting, searching), measure time saved and quality on a pilot perimeter, and keep the cases that touch citizens' rights (benefits, audits, decisions) for a second phase, with impact assessment and reinforced human oversight. That is the path that minimises legal risk and maximises learning.
List repetitive tasks, volumes, available data and their quality. The use-case inventory, as practised by US agencies, is also a transparency tool.
One or two low-risk internal cases, a test group, a defined duration and indicators measured before and after: processing time, quality, staff satisfaction.
Usage charter, impact assessment (DPIA), AI Act classification, logging and human oversight. The Belgian federal charter of 11 July 2025 offers a directly reusable reference framework.
Generalise what proved its value, train staff, publish what you automate and why. Citizens' trust is earned through transparency, not fait accompli.
The public sector is on the front line of the EU AI Act. Prohibited practices (social scoring, real-time remote biometric identification save for strict exceptions) have applied since 2 February 2025. Systems that condition access to essential public benefits and services (allowances, aid, emergency services) fall under high risk: risk management, data quality, technical documentation, human oversight and registration in the European database. After the postponement decided in spring 2026, these obligations apply on 2 December 2027: the delay is for preparing, not waiting.
The GDPR remains the foundation: legal basis, minimisation, impact assessment and strict framing of fully automated decisions (Article 22). Add the hosting question: for citizens' data, the choice of provider and hosting region is as much a matter of digital sovereignty as of technology. The Union is also pushing for action: the Apply AI strategy, published on 8 October 2025, makes the public sector one of its 11 priority sectors and announces an AI observatory for 2026.
With an inventory of its possible use cases and data, then with one or two high-volume, low-risk internal cases: summarising documents, preparing replies, searching knowledge bases. Measure time saved and quality, frame with a charter and a risk analysis, then expand. Cases touching citizens' rights come last, with reinforced safeguards.
No, not for decisions that produce legal effects on a person. Article 22 of the GDPR strictly frames fully automated decisions, and the EU AI Act classifies systems used for access to essential public services as high risk, with mandatory human oversight. AI prepares, sorts and suggests; the public official decides and remains accountable.
Prohibited practices (social scoring, real-time remote biometric identification save for exceptions) have applied since 2 February 2025. Systems that condition access to essential public benefits and services are classified as high risk: risk management, data quality, documentation, human oversight. After the postponement decided in spring 2026, these obligations apply on 2 December 2027.
A clear legal basis, data minimisation, an impact assessment (DPIA) before any risky processing, and no citizens' data sent to an uncontrolled consumer AI tool. Hosting and data flows must be documented, with a preference for European solutions or solid contractual guarantees, and explicit transparency whenever a citizen interacts with an AI.
Molderez Consult helps Belgian administrations, public and semi-public bodies inventory their use cases, launch measured pilots, frame the risks (EU AI Act, GDPR, usage charter) and train staff, from the first internal assistant to citizen-facing services.
Discuss my project