The UK's DWP fraud model: a machine-learning system that saved an estimated 4.4 million pounds on Universal Credit advances

The problem

Universal Credit advances are paid quickly, which is exactly what makes them a target for fraudulent claims. Screening every advance by hand is not feasible at the scale the Department for Work and Pensions operates, so DWP turned to data analytics to decide which claims warrant a closer look.

What was built

Per the National Audit Office, “since May 2022, DWP has used a machine learning model to flag potentially fraudulent claims for Universal Credit advances” (source). The model scores claims for fraud risk and routes the flagged ones to human reviewers rather than deciding outcomes on its own.

The outcome

An estimated 4.4 million pounds saved. Per the NAO, the model has been “saving an estimated £4.4 million” (source).

Around three times more effective than random sampling. “DWP found the machine learning model to be around three times more effective at identifying fraud risk than a randomised control group sample” (source). The comparison against a randomised control group is what makes this figure more than an assertion, though the underlying numbers are still DWP’s own estimates.

The counter-context. The same NAO release notes that “concerns have been raised by the Public Accounts Committee about the potential impact of machine learning on vulnerable claimants” (source). This file reports the model’s stated results and that caveat together.

How this was verified

Every figure on this page was re-checked, word for word, against the live NAO release on 2026-08-15, and each held. Its footing is better than a vendor case study: the figures reach the public through the National Audit Office, the UK’s independent public-spending auditor, and rest on a comparison against a randomised control group. But they are still DWP estimates, single-sourced, and not independently re-measured, and the Public Accounts Committee has flagged the potential impact of machine learning on vulnerable claimants. TIN is an independent audit of the public record, not an outreach to the subject: the badge never depends on DWP confirming anything, and the NAO here is only relaying DWP’s own estimates rather than independently measuring them. That single-source, DWP-originated limit is the honest ceiling on these numbers and is stated here so the reader meets it alongside them.

Related case files

The pattern of an automated model scoring individuals, then a public body questioning its fairness, runs through several files here. In Louis v. SafeRent, an algorithmic tenant-screening score was rolled back for voucher applicants under a class settlement, the same tension between an efficiency claim and its impact on the people being scored that the Public Accounts Committee raises here. In EEOC v. iTutorGroup, automated screening that rejected applicants by age drew a federal consent decree, another case of an automated gate weighed against a protected group. And the Dutch DPA’s fine on Clearview AI shows an independent regulator, not the deploying body, setting the record on an AI system, the same reason the NAO’s relay matters more than a DWP press release would.

Sources

Cited in this case file. Tier 1 = independent public auditor relaying government figures. Each figure was checked against the live source on 2026-07-10 and re-checked, word for word, on 2026-08-15.

1. National Audit Office, “DWP begins to make headway tackling benefit fraud and error” (Tier 1, independent auditor; figures are DWP estimates). https://www.nao.org.uk/press-releases/dwp-begins-to-make-headway-tackling-benefit-fraud-and-error/ · archived