Machine learning saves £4.4M in UK.gov work and pensions fraud detection

TL;DR: The UK’s Department for Work and Pensions achieved £4.4 million in savings over three years by deploying machine learning to combat fraud, according to the National Audit Office. However, fragmented IT systems limit scaling efforts, and fairness concerns persist with applicants aged 45+ and non-UK nationals facing higher flagging rates despite lower claim refusal rates.

The UK government’s Department for Work and Pensions (DWP) has demonstrated measurable results from machine learning deployment in fraud detection, though significant technological and fairness challenges constrain expansion.

Context and Background

In its October 22 report on tackling benefit overpayments, the NAO acknowledged DWP’s progress whilst recommending further advancement. Report director Laura Brackwell stated: “DWP should build on its existing use of data analytics to explore how these emerging technologies may help detect and prevent fraud and error.”

The context reveals the scale involved: DWP distributed £291 billion to 23 million people in 2024-25—exceeding healthcare spending and tripling defence expenditure. The £4.4 million savings calculates to approximately two pence per UK resident annually, demonstrating both the achievement and the opportunity for expansion.

Technical Limitations and Scaling Challenges

The core challenge stems from technological limitations. DWP’s IT infrastructure lacks full integration, preventing staff from accessing comprehensive claimant information. The department is developing an application to provide unified views of claimant data, though scaling requires establishing cross-government data standards for inter-departmental information sharing.

Denmark provides a relevant comparison, having implemented interoperable IT systems and government-wide data standards that support approximately 100 anti-fraud machine learning models—suggesting the UK’s five operational and development-stage models represent only a fraction of what integrated infrastructure could enable.

Current Implementation and Fairness Concerns

DWP’s current machine learning initiative focuses on Universal Credit, which consolidates multiple legacy benefits. Since May 2022, a deployed model has flagged potentially fraudulent hardship payment advance claims for human review rather than automatic denial.

The NAO identified fairness concerns in model performance. Applicants aged 45 and older, plus non-UK nationals, faced higher flagging rates but lower claim refusal rates. The department assessed only one of nine legally protected characteristics—age—citing insufficient data on others.

Despite these limitations, the model demonstrates three times greater effectiveness than random sampling and remains operational pending improvements. This pragmatic approach—deploying imperfect systems whilst working towards improvement—reflects the tension between immediate fraud prevention and fairness requirements.

Looking Forward

Four additional machine learning models are currently under development, all targeting Universal Credit. These address undeclared self-employment income, financial assets, undisclosed partners, and general fraud detection patterns.

The NAO recommended standardizing claimant data formats, engaging with cross-government data initiatives, and extending anti-fraud efforts to other benefits, particularly Pension Credit, which demonstrated the highest overpayment rate in 2024-25.

For public sector organizations considering machine learning deployment, DWP’s experience demonstrates both opportunity and constraint. The £4.4 million savings prove value, but fragmented IT infrastructure, incomplete protected characteristic data, and fairness assessment challenges limit scaling potential. Denmark’s more developed infrastructure suggests that integrated systems and data standards represent prerequisites for substantial expansion rather than optional enhancements.

Source Attribution:

Share this article