For decades, banking fraud detection relied on hand-coded rules. If a transaction exceeds a certain amount in an unusual country, block it. If three attempts fail in under a minute, trigger an alert. These rules are predictable, and therefore bypassable.
What the agentic approach changes
Agentic systems do not follow static rules. They analyse each transaction in context: the customer’s past behaviour, geolocation, merchant type, time of day, access channel, device used. They generate a contextual decision, not a rule application.
The practical difference is significant. An unusual transfer made from the customer’s regular mobile app, at their registered home address, from their known device, will be treated differently from a transfer of the same amount made via an unknown browser while roaming.
What Société Générale measured
In 2025, Société Générale published partial data on its deployment. The false positive rate (legitimate transactions wrongly blocked) fell by 31% after switching to an agentic system, across a scope covering retail debit cards.
For customers, this means fewer legitimate transactions blocked during travel or unusual purchases. For the bank, it means lower costs in processing complaints.
The limits of the system
AI agents can also be circumvented, just differently. Fraudsters adapt their behaviour to mimic legitimate patterns. An agent that learns continuously can be manipulated through gradual injection of fraudulent behaviours that resemble a slow drift in legitimate usage.
Fraud prevention teams describe a cat-and-mouse game that has accelerated, not been resolved.
Key takeaway
AI agents have substantially reduced false positives and improved detection of complex fraud. They have not ended fraud. They have displaced and accelerated the adaptation of fraudulent techniques.