AI in government examples: what the audited record shows (2026)

Search for AI in government examples and you get lists of use cases: chatbots for permits, models for fraud, forecasting for budgets. Almost none of them says what happened after the press release.

That gap matters more in government than anywhere else. A public agency cannot quietly retire a bad system. Its failures land on benefit claimants, and its successes are claimed in budget statements.

The scale of the bet is now explicit. The UK government expects “substantial efficiencies from digital transformation and AI, amounting to £45 billion each year,” according to the National Audit Office [source].

So the useful question is not what governments could automate. It is which examples have an auditor, a court or a regulator on the record, and what those bodies found.

What AI in government means

AI in government is the use of software that scores, flags or decides cases inside a public agency, from fraud models to automated benefit decisions. In practice the label covers rule engines as often as machine learning.

That second sentence is the first thing the record teaches. Two of the six cases below were never machine learning at all, and the harm was the same.

What are the documented AI in government examples?

Six AI in government examples have an independent body’s findings on the public record and a TIN case file behind them. One is a measured success, one is an unevidenced claim, and four were found wrong or unlawful.

UK DWP fraud model. “Since May 2022, DWP has used a machine learning model to flag potentially fraudulent claims for Universal Credit advances,” the NAO reports, with flagged claims going to human reviewers [source]. The full case is in the DWP case file.

UK £45 billion efficiency claim. The NAO found “the published efficiency plans do not provide details of how departments derived their expected workforce efficiencies” [source]. See the NAO workforce case file.

US VA survivors’ benefits. The VA inspector general found “at least 8,000 decisions or letters omitted favorable findings” in automated death-benefit claims [source]. See the VA OIG case file.

Michigan MiDAS. Of the penalty cases the state later re-examined, 40,195 had been resolved by computer and 85 percent of those fraud findings were reversed [source]. See the MiDAS case file.

Dutch SyRI. On 5 February 2020 the District Court of The Hague found the SyRI welfare-fraud legislation breached Article 8 of the European Convention on Human Rights [source]. See the SyRI case file.

Dutch childcare benefits. The Dutch DPA fined the Tax Administration €2.75 million for using nationality as an indicator in a system that “automatically designated certain applications as risky” [source]. See the toeslagenaffaire case file.

The record in one table

Every figure in the table is sourced in the section above or below it. The table is a map of who checked, not a score of the technology.

What are the benefits of AI in government?

The benefits of AI in government that someone outside the agency has examined are real but small. The DWP model is “saving an estimated £4.4 million,” the NAO reports [source].

The stronger number is the comparison. “DWP found the machine learning model to be around three times more effective at identifying fraud risk than a randomised control group sample” [source]. A control group is what turns a vendor claim into a measurement.

The limits sit in the same release. Both figures are DWP’s own estimates relayed by the NAO, and “concerns have been raised by the Public Accounts Committee about the potential impact of machine learning on vulnerable claimants” [source].

Set that £4.4 million beside the £45 billion a year the government expects [source]. The measured benefit and the planned benefit are not yet on the same scale, and the NAO could not find the working for the larger one.

Where AI in government went wrong

Government automation failed in this record where software decided a person’s case with no human review. The VA inspector general said its errors “will recur because the system applies predefined rules without human involvement” [source].

The money followed the errors. “At least 2 percent of decisions had legal errors, resulting in an estimated $2.7 million in improper payments,” the inspector general found [source].

Michigan shows the human step mattering inside one system. Cases MiDAS resolved by computer alone were reversed 85 percent of the time; the 22,589 cases with some human involvement had a 44 percent false fraud rate [source].

The human step cut the error rate without fixing it. The Michigan Supreme Court later recorded that “approximately 93% of the automated system’s fraud determinations were incorrect” [source]. In January 2024 more than 3,000 people falsely accused were in line for a share of a $20 million settlement ([source]).

The Dutch cases failed on a different axis: transparency and discrimination, not accuracy. The Hague court found “the application of SyRI is insufficiently transparent and verifiable” and declared the legislation of no binding effect [source].

The childcare-benefits case cost more than a fine. A parliamentary inquiry found approximately 10,000 families had been forced to repay subsidies, and the whole Dutch government resigned on 15 January 2021 [source].

Is it even AI?

Not always, and the record is explicit about it. FedWeek reported the VA system is “not specifically characterized as artificial intelligence” and “applies predefined rules to generate decisions” [source]. MiDAS was an automated rules system too.

The obvious reading of this list is that government AI fails. That reading is wrong. The failures did not depend on the model being clever or crude; a fixed rule caused as much damage as a learned score.

What separated the one positive example from the rest was design, not technology. DWP’s model flagged claims for a human to decide, and DWP measured it against a randomised control group [source]. None of the failures was built to be measured that way before it went live.

The proof

Each example above rests on a TIN case file that checked its figures against the public record: audit reports, court judgments and regulator decisions, quoted as written. The case files carry the method and the sources; this post only argues from them.

Read the DWP fraud model and the VA inspector general review side by side. Same kind of agency, same kind of benefit, opposite results, and the difference is the human reviewer and the control group.

The Dutch and Michigan cases also appear in the automated decision making post, which reads them as a GDPR and due-process question. This post reads them as a buyer’s question: what did the public sector actually get?

The bottom line

The best AI in government example on the record is a £4.4 million saving, and the biggest is a £45 billion expectation nobody has shown the working for. A public buyer should treat any use case as a claim until someone outside the agency has measured it against a control group, with a human deciding the case.

That rule applies far beyond government: an example is only as good as whoever measured it.