Beyond The Pilot: Enterprise AI in Action Mastercard's 160 Billion Transactions: AI's Biggest Test
19 snips
Feb 4, 2026 Chris Mertz, SVP of Data Science at Mastercard who builds large-scale ML systems, and Johan Gerber, EVP of Security Solutions with a law-enforcement background, pull back the curtain on running AI at massive scale. They discuss scoring 160 billion transactions under a 50ms limit, an RNN “inverse recommender,” GenAI honeypots that bait scammers, and org and cloud choices that made it all possible.
AI Snips
Chapters
Transcript
Episode notes
50 Millisecond Inference At 70k TPS
- Mastercard processes ~160 billion transactions a year and the AI has ~50 milliseconds to score each transaction within a ~300ms end-to-end decision window.
- Peak throughput hits ~70,000 transactions per second during events like Black Friday and Dec 23, forcing extreme low-latency engineering.
Gas Station ID Change Broke Profiles
- A gas-station acquirer changed station IDs and Mastercard's profiles "went haywire," illustrating fragility of features tied to external identifiers.
- The incident motivated building resiliency so model profiles don't break when merchants change IDs.
Inverse Recommender RNN For Fraud
- DI Pro uses a recurrent neural network treated as an "inverse recommender" that predicts whether the current merchant is reachable from a user's past merchant sequence.
- The RNN evaluates merchant-to-merchant connections (shared customers) and treats fraud detection as pattern completion rather than anomaly detection.