Stripe Radar Adds Generative AI Rules Engine: Merchants Describe Fraud Patterns in Plain English

Stripe launched a generative AI layer for its Radar fraud prevention product that allows merchants to create custom detection rules by describing suspicious patterns in plain English. Instead of writing boolean logic like "block if (country != billing_country AND amount > 500)," merchants can now type "block transactions where the shipping country doesn't match the billing country and the order is over $500."

The AI layer translates natural language descriptions into Radar rules, tests them against historical transaction data, and shows projected impact (fraud caught vs. legitimate orders blocked) before activation. Merchants can iterate conversationally, asking the system to "make it less aggressive" or "also catch cases where the email was created in the last 24 hours."

In beta testing across 400 merchants, the natural language interface reduced rule creation time from an average of 45 minutes to 3 minutes and increased rule adoption by 4x — because fraud analysts no longer needed engineering support to implement their intuitions.

"Fraud knowledge lives in the heads of analysts, not in code," said Will Cashman, Stripe's head of Radar. "The bottleneck was always translating analyst intuition into rule syntax. We eliminated that bottleneck entirely."

The feature is available to all Stripe Radar users at no additional cost. Competing fraud platforms from Forter, Riskified, and Signifyd have not yet announced similar natural language interfaces, though several have demonstrated prototype capabilities at industry conferences.