A model reads the transaction, the device and session, the customer’s own spending history and the relationships between accounts, then returns a probability that the payment is fraudulent. It calculates that score in milliseconds, before authorisation completes. A policy engine converts the score into an approve, decline, step-up or review, applying your mandatory rules and segment thresholds on top. Confirmed fraud and chargeback outcomes flow back into training so the model keeps pace with new patterns.