.png)
Mastercard
Finance · US
200% better fraud detection
Mastercard's transformer-based AI scores 1 trillion+ data points per decision in under 50 milliseconds
Decision Intelligence Pro applies the same transformer architecture used in large language models, but to payment transaction sequences. It identifies compromised cards, merchants, and ATMs dramatically earlier than the previous generation of fraud models.
Friction
Global payments fraud costs $500B+ per year. Mastercard's existing ML models were effective but could not keep up with increasingly sophisticated criminal pattern evolution, particularly in card-not-present e-commerce fraud.
Breakthrough
Transformer models learn patterns from transaction sequences including acquirer, merchant, amount, time, and location, encoding each transaction as an embedding. At every new transaction, the full recent sequence is passed to the model. A merchant knowledge graph of billions of entities provides additional context.
Impact
200% improvement in detecting compromised cards before fraud occurs. 300% improvement in identifying compromised merchants. 85% improvement in detecting compromised ATMs. Average large issuer saves $20M+ per year in fraud losses. False-positive ratio significantly reduced, meaning fewer legitimate transactions blocked.
Unlock the full analysis with breakthrough, impact, what made it smart and its technical approach bellow!
Problem
Credit card fraud is an arms race: as detection systems improve, criminals adapt their patterns. Traditional ML models trained on static feature sets cannot track this evolution fast enough. New attack patterns require a system that can learn sequences, not just features.
What made it smart
Applying transformer architecture to transaction sequences is the conceptual breakthrough. Just as LLMs understand language by learning the relationships between words in sequence, Decision Intelligence Pro understands fraud by learning the relationships between transactions in sequence, including merchant, location, time, and behavioural context.
Technical approach
Transformer models process the full transaction sequence history for each card. At decision time, the model receives the current transaction plus recent history and generates a fraud probability score in under 50 milliseconds. A continuously updated merchant knowledge graph adds contextual signals about merchant behaviour patterns. Bank rules serve as the final filter.
Strategic lesson
The architectures that revolutionised language AI can be retrained on any sequential data, whether transactions, sensor readings, or process logs. The underlying insight is transferable even if the domain is completely different.
Reflection question
Where in your operations are you trying to detect patterns in sequences of events, and are you currently treating each event in isolation rather than as part of a chain?