Mastercard's Decision Intelligence processes 125 billion transactions per year in 50 milliseconds, setting the standard for AI fraud detection at global scale
Mastercard's Decision Intelligence platform — powered by recurrent neural networks and generative AI (Decision Intelligence Pro) — assesses fraud risk across 125 billion transactions per year in 50 milliseconds per transaction, evaluating 500+ data points to detect anomalies at the scale of a global card network.
Background
With hundreds of billions of transactions crossing its network annually, Mastercard needed a fraud detection system that could operate in near-real-time across every transaction globally — preventing fraud losses while minimizing false declines that block legitimate purchases.
What Was Implemented
- Decision Intelligence: ML-powered fraud scoring evaluating 500+ data points per transaction in 50 milliseconds, using recurrent neural networks trained on 125B annual transactions
- Decision Intelligence Pro (2024): generative AI upgrade using transformer models to model cardholder historical merchant visit patterns and generate fraud probability pathway scores
Results
The system processes 125 billion transactions per year at 50 milliseconds per transaction. Specific fraud-rate reduction percentages were not disclosed in public Mastercard sources fetched. Note: the book's "1.9 million transactions per hour" appears to materially understate Mastercard's current volumes (~14.3M/hr based on stated annual figures).
Lessons
- Network-scale training on the full cross-institution transaction graph provides detection advantages that isolated per-bank models cannot replicate
- Generative AI (transformer architectures) adds value in fraud detection by modeling sequences of merchant visits, not just point-in-time transaction features
- Processing speed (50ms per transaction) is as critical as accuracy — fraud decisions must complete before the transaction clears