Banks Turn to AI That Spots Fraud Gangs by Their Patterns

Banks are turning to AI that looks for the patterns fraud gangs leave behind, instead of relying only on fixed rules. According to Paul Chang, who heads Payment Networks at AWS (Amazon's cloud business), criminals now use packaged software that changes faster than a bank can rewrite its rules.

What changed

Chang told PYMNTS that fraudsters are using ready-made software toolkits sold on dark web marketplaces. These help them run scams, card testing attacks and synthetic identity fraud at large scale. Card testing means trying small charges to see which stolen card numbers work. Synthetic identity fraud means building a fake person from mixed real and made-up details.

Older rule-based systems are not enough against this, Chang said. Banks are moving to AI that scores transactions in real time and keeps retraining as attacks shift.

How the new tools work

One new tool is the transaction foundation model. Instead of judging a single payment alone, it learns from long sequences of financial activity. Chang described these as models trained on billions of financial events. They are meant to add to the systems banks already run, not replace them.

The second tool targets fraud rings. Their activity is spread across accounts, identities, IP addresses and devices. Graph neural networks treat accounts, transactions and devices as dots and study how they connect. A shared device or IP address can expose activity that another model would miss.

Chang said adding these layers has consistently improved detection. He pointed to Stripe's Radar, which reduces fraud by 32% on average. He also said Razorpay found five times more fraudulent or disputed transactions without adding alerts.

Where people stay in charge

Finding fraud is only part of the job. Banks must also investigate alerts and gather evidence. AI agents are starting to help by pulling records, such as device data and customer profiles, and drafting summaries with recommendations.

Chang was clear about the limit. Human analysts keep final authority over case decisions and suspicious activity report filings, known as SARs. In his words, the human is still in the loop.

What is not known yet

The article does not name which banks use these tools or when customers might notice a difference. Chang also said banks need unified, well-governed data before any of this works well, and many hold information in separate systems. For customers, the practical point is simple: faster detection helps, but a person still decides cases, so report suspicious charges quickly.