Small Business Administration piloted AI to find pandemic loan fraud without required safeguards

The US Small Business Administration (SBA) has been using an AI tool to detect fraud in its pandemic loan programmes without classifying it as a high-impact use, which would have required minimum risk management practices, its Office of Inspector General found in an advisory reported by Nextgov/FCW.

The pilot looked for fraud in disaster assistance loans, including the Paycheck Protection Program and Economic Injury Disaster Loans. It was not listed in the agency's public AI use case inventory, the watchdog said, although SBA had been open about using it.

A tool with consequences

In February SBA said it had suspended 111,620 California borrowers suspected of pandemic-era fraud and that it had "recently partnered with Palantir to expand the nationwide investigation". In July it formalised its use of Palantir software for fraud prevention.

Under the Office of Management and Budget's April 2025 memo on federal AI, systems that affect people's rights or access to services are presumed high-impact. They must be tested before deployment, assessed for impact, monitored and subject to human review, and agencies must offer a way to appeal.

Why classification matters

Fraud detection is one of the strongest uses of AI in government, and the pandemic loan programmes lost large sums to fraud. But flagging a borrower can lead to suspension, debt collection or prosecution, and false positives fall on real businesses. Labelling the tool high-impact is not paperwork: it is what triggers the testing and appeal rights that protect the people a model wrongly flags.