BetaThis is a new service. Your feedback will help us improve it. Tell us what you think

AI-Driven Counter Fraud System

The counter fraud system helps investigators prioritise which benefit applications to examine, by looking for patterns associated with previously confirmed organised fraud.

Keywords

  • Tags: Fraud detection
  • Tags: Prioritisation
  • Tags: Risk scoring
  • Stage: Pilot
  • Type: Gradient-boosted decision trees
  • Sector: Social security
  • Language: English

How does our product work?

The system reads an application after it has been submitted and compares its features — not the applicant's characteristics — against patterns found in confirmed fraud cases. It returns a score and the three features that contributed most to it.

A high score routes the case to a human investigator sooner. It has no effect on entitlement, and it cannot delay or refuse a payment on its own.

Overview

Organised benefit fraud is a small share of claims but a large share of loss, and it moves faster than manual sampling can follow. This system exists so that investigator time goes where it is most likely to matter.

Every output is advisory. The decision to open an investigation, and the decision on the claim itself, remain with named officers.

Owner and responsibility

Social Security Agency

Ada Whitfield

counter-fraud@example.gov

See more information

More detailed information on the system

Here you can get acquainted with the information used by the system, the operating logic, and its governance in the areas that interest you.

System description

Features are drawn from the application itself and from patterns across applications, such as repeated bank details or addresses. Protected characteristics are excluded from the feature set and are used only afterwards, in aggregate, to test the model for disparate impact.

The model is retrained quarterly. Each retraining is accompanied by a fairness assessment across age, sex, disability status and postcode deprivation decile, published in the annual report.

Data sources

Benefit application records. Confirmed fraud case outcomes from the previous six years. No data is purchased from third parties and no social media data is used.

References

Fairness assessment, 2026 edition.

Pilot evaluation covering the first nine months of operation.

Was this information useful?

Would you like to give feedback? Your feedback will help us develop our algorithms further.

If you wish for the system owner to get back to you personally, please contact them directly using the email address above. For more general comments, use the box below.