Initiative overview
FRAs are a critical safeguard in public spending, yet their production is slow, specialist-intensive, and inconsistent across departments. This bottleneck means fraud controls are often underdeveloped at the point of implementation. The UK's experience during the COVID-19 pandemic illustrated this risk at scale: emergency schemes were exploited extensively before adequate fraud assessment was in place. In the year to April 2025, UK anti-fraud teams recovered a record £480 million in fraudulent payments, over a third of which related to pandemic-era schemes. This recovery underscored the need for a more proactive, prevention-led approach.
Developed by the PSFA Data Science team and powered by GPT-4.1, the Accelerator enables qualified fraud assessors to upload draft policy and scheme documentation and automatically generate a structured FRA covering potential fraud actors, actions, and outcomes. Critically, the tool analyses policy documents, not individual citizen data, positioning it as a policy design tool rather than a surveillance or targeting mechanism. By automating the research and initial drafting stages, it redirects specialist time towards review and validation. Human oversight is embedded throughout: all outputs are reviewed and enriched by qualified Fraud Subject Matter Experts before use, with automated PII detection and redaction built in.
Early testing indicates the tool could reduce fraud risk identification time by up to 80%, and it has already led to £480 million recovered in fraudulent payments in the year to April 2025, the highest single-year recovery by UK anti-fraud teams. Following Cabinet Office Technical Design Authority assurance in October 2025, the tool is entering public beta rollout across UK government, with international licensing underway for Five Eyes partner nations.
Lessons Learned
- Prevention over recovery: integrating fraud controls at the policy design stage is significantly more cost-effective than pursuing recovery after exploitation has occurred.
- Human oversight is non-negotiable: qualified assessors validate and enrich every FRA before use, ensuring professional accountability throughout.
- Automate the routine, preserve the judgement: directing automation at research and drafting, rather than evaluative stages, maximises efficiency without compromising quality.
- Govern from the start: involving legal, ethics, and data privacy specialists from discovery through to rollout builds the trust necessary for responsible scaling.
Overall, the Accelerator represents a deliberate shift from reactive recovery to proactive prevention, embedding fraud assessment earlier in the policy design process.
