Fraud, risk & detection

Real-time payment fraud detection

Source Mastercard (Decision Intelligence / Decision Intelligence Pro) — public case. This is an industry example, not our project

How does AI detect payment fraud in real time?

Payment fraud has to be caught in the moment — in the milliseconds between a card being tapped and a transaction clearing — without blocking so many legitimate payments that customers give up. One payments network uses AI to score every transaction against learned patterns in real time, catching more fraud while cutting false declines. The transferable capability is sub-second anomaly detection on high-volume transaction streams, tuned to the trade-off between catching fraud and waving through genuine customers.

~20%
average improvement in fraud detection
>85%
reduction in false positives (own analysis)

The problem

Card fraud is a vast, adaptive, multi-trillion-dollar problem; rule-based systems can’t keep pace with evolving patterns and generate costly false declines.

The AI approach

Decision Intelligence scores transactions in real time. The 2024 “Pro” upgrade uses a proprietary generative-AI/transformer model (with graph machine learning) trained on ~125 billion annual transactions, assessing the relationships between entities to judge whether a transaction fits a cardholder’s behaviour — in ~50 milliseconds.

Evidence it works

Mastercard reports average fraud-detection improvements of ~20% (and up to 300% in some cases), false-positive reductions of over 85% in its own analysis, and — via a separate 2024 generative-AI tool — doubling the rate of compromised-card detection.

What “good” looks like

Higher fraud catch rates and fewer false declines (the false-positive number matters as much as the catch rate), with sub-100ms scoring at network scale.

Feasibility & cost shape

Network-scale capability built on enormous data and sustained investment (Mastercard cites $7bn+ over five years); most institutions consume it as a service.

Our independent view

A strong example of why data scale and continuous learning beat static rules in adversarial domains. We’d judge any such tool on the false-positive trade-off, not the catch rate alone.

Source & attribution

Based on publicly reported information about the Mastercard (Decision Intelligence / Decision Intelligence Pro) work.

This is an industry example included for illustration. It is not a Leia Intelligence project, and no client of ours is implied. Figures are as publicly reported by the original parties.

Sources: CNBC · PYMNTS · Mastercard press releases (Feb & May 2024) · TechInformed

This is the class of problem we help organisations tackle — book a call.
Book a call