Anti-fraud with proprietary models: real case in payments
We reduced chargebacks by 68% without adding friction. How we combined graph analysis, multivariate anomaly and LLMs for assisted investigation.
The client is a payments processor with hundreds of millions of monthly transactions. The operation depended on an external scoring vendor, with high latency, growing false positives and zero explainability. We left with a proprietary model in production in 14 weeks.
The starting point
Three bad metrics at once: chargebacks at 0.38% of volume, false positives at 12% (legitimate customer blocked), average scoring latency at 220ms. The scoring came from a third-party API with opaque rules and a high unit cost.
What we changed
We built three layered models: a fast classifier for 95% of cases, a graph model to detect mule networks, and a multivariate anomaly model to detect atypical behavior. On top, an LLM agent that prepares the case for the human investigator with hypotheses already organized.
The results, 90 days later
Chargeback fell to 0.12% (-68%). False positives fell to 5.5% (-54%). Average latency settled at 78ms — three times faster than the vendor. The investigation team now resolves 3.2x more cases per day, because the agent has already done the context-gathering.