AI in Government: How Public Services Are Being Transformed by Machine Learning

How governments use AI in 2026: 2,133 U.S. federal use cases, what OMB M-25-21, NIST, the EU AI Act, and Canada's directive require, lessons from the Dutch benefits scandal, and a deployment checklist.

PUBLISHED // UPDATED
DURATION // 09 MIN READ
AI in Government: How Public Services Are Being Transformed by Machine Learning
FIG.01 // VISUAL DOCUMENTATION AI in Government: How Public Services Are Being Transformed by Machine Learning

Public-sector AI rarely looks like the futuristic systems in vendor demos. Most of it is administrative: sorting correspondence, pre-filling forms, routing cases, detecting anomalies in claims, and scheduling maintenance. These uses can save time for citizens and staff, but government AI also carries risks that private-sector tools usually do not. A wrong decision can deny someone benefits, flag them as a fraudster, or affect their legal status, and citizens usually cannot choose another provider.

This guide explains where AI is being used in government, what published frameworks require, what past failures teach, and how public-sector teams can deploy AI responsibly.

How much AI government already uses

The clearest public data comes from the United States, where federal agencies must publish inventories of their AI use cases. According to the consolidated 2024 Federal AI Use Case Inventory maintained by the Office of Management and Budget, as of January 23, 2025 there were 41 agency submissions covering 2,133 AI use cases, of which 351 were classified as rights-impacting and/or safety-impacting.

Two lessons follow from those numbers. First, AI is already widespread across government, not just in a few pilot projects. Second, most uses are not classified as affecting rights or safety, which reflects the fact that much government AI supports internal processes rather than making decisions about individuals. The smaller group of high-stakes uses is where oversight matters most.

Where AI tends to work well

Administrative processing and routing

Benefits agencies, tax authorities, and licensing offices receive large volumes of forms and documents. AI can classify incoming documents, extract information, detect missing items, and route cases to the right team. Done well, this shortens queues without letting software make the final decision. The key design choice is that the system prepares a case for a human decision-maker instead of replacing that person.

Shared digital infrastructure

Estonia is often cited as an example of digital government, and its foundation is less about AI than about data infrastructure. Its X-Road data exchange layer connects public and private databases so that agencies can securely request information from one another rather than asking citizens to submit the same documents repeatedly. X-Road is open-source software and is also used outside Estonia.

The relevance for AI is direct: automated pre-filling, eligibility checks, and anomaly detection all depend on accurate, connected, and well-governed data. Governments that try to add AI on top of fragmented records usually get unreliable results.

Maintenance and operations

Transportation agencies and utilities increasingly use sensor data and machine learning to prioritize inspections and maintenance of roads, bridges, water networks, and vehicle fleets. These uses are generally lower risk for individual rights because they concern physical assets, although errors can still have safety consequences. They are often a sensible place to build skills and data practices before applying AI to decisions about people.

Citizen-facing assistants

Chatbots and virtual assistants can answer common questions and help people find the right service. They are useful for simple, well-documented information, but they must not give confident wrong answers about eligibility or legal rights. A good assistant states its limits, links to authoritative pages, and offers a clear route to a human.

What the frameworks require

Several published frameworks shape how governments deploy AI. They are worth reading directly rather than through summaries.

FrameworkJurisdictionKey requirements
OMB Memorandum M-25-21U.S. federal agenciesEach agency identifies a Chief AI Officer; minimum risk-management practices for “high-impact AI”
NIST AI Risk Management FrameworkU.S. (voluntary)Four functions: Govern, Map, Measure, Manage
EU AI Act (Regulation 2024/1689)European UnionRisk-based rules; many public-benefit uses are high-risk; fundamental-rights impact assessments for public bodies
Directive on Automated Decision-MakingGovernment of CanadaMandatory Algorithmic Impact Assessment; requirements scale with impact level
Algorithmic Transparency Recording StandardUnited KingdomPublic records describing algorithmic tools used by public bodies

United States

OMB’s April 2025 memorandum M-25-21 requires agencies to designate a Chief AI Officer and to put risk-management processes in place for high-impact AI, meaning AI whose outputs serve as a principal basis for decisions with significant effects on people’s rights or safety. The NIST AI RMF, released in January 2023, is voluntary but widely used as a common vocabulary for identifying and managing AI risks, and NIST has added a profile for generative AI (NIST AI 600-1).

European Union

The EU AI Act entered into force on August 1, 2024, with obligations phased in over several years. Among other things, it treats AI used to evaluate eligibility for essential public benefits and services as high-risk, which brings requirements for risk management, data governance, documentation, human oversight, and accuracy. Public bodies deploying such systems must also carry out a fundamental-rights impact assessment.

Canada and the United Kingdom

Canada’s Directive on Automated Decision-Making requires federal institutions to complete an Algorithmic Impact Assessment before using automated decision systems. The assessment assigns an impact level, and higher levels require more safeguards such as peer review, human intervention, and explanations. The UK’s Algorithmic Transparency Recording Standard asks public-sector organizations to publish structured descriptions of the algorithmic tools they use, including their purpose and how they fit into decisions.

What past failures teach

The Dutch childcare benefits scandal

In the Netherlands, the tax authority’s handling of childcare benefit fraud led to thousands of families being wrongly accused of fraud and ordered to repay large sums. The Dutch government resigned in January 2021 over the scandal. Amnesty International’s report Xenophobic Machines examined the role of an algorithmic risk-classification model in the system and argued that its design and use, including the use of nationality as a risk factor, led to discrimination.

The lessons are not only about the algorithm. They include the lack of transparency, the difficulty families had in contesting decisions, and the way automated risk scores were treated as grounds for harsh action.

The 2020 UK exam grading algorithm

When exams were cancelled during the pandemic in 2020, England’s exam regulator used a statistical model to standardize grades. Results were widely criticized for downgrading many students, particularly where their school’s historical performance was weaker, and the government reversed course in favor of teacher-assessed grades. The episode showed that a technically consistent method can still be unacceptable if it produces outcomes that individuals see as unfair and cannot challenge.

Common patterns

Across failures, the same issues recur:

  • Automated outputs treated as decisions rather than as inputs for a human.
  • Little transparency about how the system worked or what data it used.
  • Weak appeal routes, so errors persisted.
  • No testing for disparate impact on particular groups before deployment.
  • Scale: a flawed system affects many people at once.

A deployment checklist for public-sector teams

Before procurement or development

  • Define the problem and the decision the system supports. Write down who makes the final decision.
  • Classify the use case: does it affect individuals’ rights, benefits, or safety? If yes, apply the stronger requirements of the relevant framework.
  • Complete an impact assessment, such as Canada’s Algorithmic Impact Assessment or an EU fundamental-rights impact assessment where required.

During design and testing

  • Test accuracy and error rates separately for relevant groups, not only on average.
  • Check data quality and whether historical data reflects past biases.
  • Design the human review step so reviewers have time, information, and authority to override the system.

At launch

  • Publish a plain-language description of the system, such as an ATRS record in the UK or an entry in a public AI inventory.
  • Tell people when automation is involved in a decision about them and how to appeal.
  • Keep audit logs of inputs, outputs, and human decisions.

After launch

  • Monitor error rates, appeals, and complaints over time.
  • Re-test for disparate impact after model or policy changes.
  • Make sure procurement contracts give the agency access to documentation and the ability to audit the system.

Questions to ask vendors

  • What data was the model trained on, and can we audit it?
  • How does the system perform for different demographic groups?
  • What explanations can it provide for individual outputs?
  • What logs are available, and who owns them?
  • How are updates handled, and will we be notified before model changes?
  • Can we exit the contract and keep our data and configuration?

For broader governance guidance, see our articles on enterprise AI governance and compliance and AI in smart cities.

Bottom line

Government AI is already widespread: U.S. federal agencies alone reported more than 2,000 use cases in their 2024 inventory. Most help with administrative work, while a smaller set affects people’s rights or safety. Frameworks such as OMB M-25-21, the NIST AI RMF, the EU AI Act, Canada’s Directive on Automated Decision-Making, and the UK’s transparency standard give agencies a clear structure for managing those risks. The failures in the Netherlands and the 2020 UK exam grading show what happens when automated outputs are treated as decisions without transparency or appeal. Public bodies that start with lower-risk uses, keep humans accountable for consequential decisions, publish what their systems do, and test for unequal impacts are far more likely to deliver AI that improves public services.

Sources