JPMorgan Deploys LLM-Based Anomaly Detection Across $10T in Daily Payment Flows

JPMorgan Chase has deployed large language model-based anomaly detection across its wholesale payments infrastructure, analyzing the natural language content of SWIFT messages, payment instructions, and beneficiary details to catch compliance and operational errors. The system processes metadata from over $10 trillion in daily payment flows.

Unlike traditional transaction monitoring (which flags numerical anomalies like unusual amounts or frequencies), the LLM-based system reads the unstructured text within payment messages — beneficiary names, purpose-of-payment fields, remittance information — and identifies semantic anomalies that rules-based systems miss.

In its first quarter of production, the system flagged 42% more suspicious activity reports than the previous rules-based approach while reducing false positives by 31%. It has also caught operational errors — such as misrouted payments and incorrect beneficiary details — before they reached clearing.

"Payment messages contain rich information that rules engines simply cannot parse," said Lori Beer, global CTO at JPMorgan. "When a beneficiary name subtly changes between related payments, or a purpose-of-payment field uses unusual phrasing, that is a signal. LLMs can read signals that pattern-matching cannot."

The deployment is the largest known application of LLMs in payment processing. Competing banks including HSBC and Standard Chartered have disclosed similar research programs but have not confirmed production deployments at this scale.