The Fraud Risk Accelerator: Using AI to Stress-Test Schemes Before Fraud Gets In

Craig Martin, Fraud Analytics Lead and Head of Programme at the PSFA, and Jennifer Evans, Product Manager, gave a live demonstration at Counter Fraud 2026 of the PSFA's Fraud Risk Accelerator - an AI-powered tool that analyses a scheme or policy document and generates a list of potential fraud vectors before a single claim has been made.

Here is a summary of what they said and what the demonstration showed:

The context: AI reasoning is catching up fast

Craig Martin opened with a chart showing how the performance gap between Large Language Models (LLMs) and human, expert-level reasoning has closed over the past year. Models that were noticeably below human expert performance in 2024 are now at or approaching this level across a range of tasks - not just information retrieval, but reasoning. That shift matters for how AI tools are being used in fraud prevention, and it is this capability underpinning the Accelerator.

What the Fraud Risk Accelerator does

The Fraud Risk Accelerator is a beta service that 33 public bodies are now using, with more joining regularly. The core function is straightforward: upload a PDF of a scheme or policy document, and the tool uses generative AI - conditioned through extensive prompting to act as an adversary - to identify where the scheme might be defrauded. The output is a structured list of potential fraud risks, categorised by actor, action and outcome.

The purpose is to support the first stage of a fraud risk assessment: identifying the risk landscape before moving on to assessing likelihood, impact and controls. It is not designed to replace the fraud risk assessor; it is designed to accelerate and enrich what they can see, surfacing risks that might not be immediately obvious from reading the scheme documentation manually.

The live demonstration: childcare grant

For the demonstration, Jennifer Evans described the scheme to be tested: the Childcare Grant for full-time higher education students with childcare responsibilities, paying up to £200 per week for one child and £340 for two or more. Eligibility criteria include full-time student status (120 or more credits), permanent residency in England, household income thresholds, and use of a registered childcare provider. Funds flow into an online account, the provider bills against it, and the parent authorises each payment.

While the tool ran - taking around three minutes - Evans asked the audience to identify their top three ways to defraud the scheme. The room came up with: falsifying student status; lying about having children or their ages; falsifying residency or using a temporary address to claim in multiple regions; false income declarations; fake or colluding childcare providers; dual claiming by both parents; using the identities of deceased children; dropping out of studies but continuing to claim; and falsifying special educational needs status to extend the eligible age.

In the same time, the tool returned 60 distinct fraud risks. Checking the audience's suggestions against the output, the accelerator had identified versions of almost every one - including the more specific and niche risks such as using temporary accommodation addresses to claim in multiple regions, failing to declare a change in student status, falsifying Ofsted registration details, both parents claiming simultaneously, and provider collusion with applicants to fabricate childcare that was never delivered. It had also identified internal fraud risks - staff within Student Finance England approving fraudulent applications in collusion - that the audience had not raised.

What the tool is and is not

Martin and Evans were consistent on a critical point: the output is a starting point, not a finished fraud risk assessment. The tool generates a broad landscape of plausible risks, covering all actors in the supply chain - applicants, providers, intermediaries and internal staff. What it does not do is apply the contextual knowledge that only the fraud risk assessor holds: the organisation's internal controls, their effectiveness, the local context, and the operational reality of how the scheme runs in practice. Those judgements - on likelihood, impact and mitigation - remain the assessor's responsibility.

Evans also noted that the 60 risks are presented in a deliberately granular, unsimplified form. The tool does not collapse similar risks together, because deciding how to group and prioritise them requires the kind of scheme-specific knowledge that belongs with the assessor, not the AI. Some risks in the list will be combined, some split further, and some deprioritised - but that decision should be made by a person.

Under the bonnet, the tool uses a multi-agent workflow with multiple LLMs chained together - one to identify risks, another to refine and quality-check them. A multi-page PDF up to 200 megabytes is well within the tool's capacity.

Where it is going next

Questions from the floor pushed on what comes next. Could the tool generate recommended controls as well as risks? Martin and Evans confirmed this is being actively explored, while noting that the human judgement involved in control design is significant - particularly because the effectiveness of controls often depends on contextual knowledge the AI does not have. One audience member asked whether an agentic AI could go further and actually attempt to defraud a scheme, testing whether it could succeed. Martin acknowledged the creative thinking, and confirmed the team would explore it with users, while noting current limitations around the model's contextual knowledge of internal processes.

A local authority representative confirmed that their organisation is part of the beta programme and is actively comparing the tool's output against fraud risk assessments completed manually - testing where the AI finds risks the human process missed, and vice versa. That kind of structured evaluation is exactly how the PSFA is learning what the tool does well and where the human remains essential.

The session's closing message was simple: the Fraud Risk Accelerator does not replace the fraud risk assessor. It gives them a richer, faster starting point - a comprehensive adversarial view of a scheme generated in minutes rather than hours, so that expert time can be focused on the judgements that only a person can make.