Danske Bank

Danske Bank's fraud detection infrastructure had grown brittle over years of reactive, rules-based policy-making. The system produced a false positive rate approaching 99.5% of all flagged transactions — meaning nearly every alert raised by the engine identified a legitimate payment rather than genuine fraud. In retail banking, where payment volumes span millions of online, card, and mobile transactions daily, this translated into an enormous and unsustainable investigation burden. A large dedicated fraud team spent the majority of its time reviewing clean transactions rather than pursuing real threats. The operational costs were significant, and the system's inability to adapt to novel fraud patterns left the bank increasingly exposed as payment channels expanded.

Working in collaboration with an AI vendor specializing in financial crime, Danske Bank replaced its legacy rule engine with a real-time machine learning fraud platform covering its full payment stack — online banking, credit card purchases, and mobile payments. The system was engineered to score each incoming transaction in under 300 milliseconds, fast enough to intervene at the point of authorization without introducing perceptible latency for customers at the point of sale. The ML engine analyzes tens of thousands of latent features per transaction, processing millions of events to surface behavioral patterns and statistical signals that no manually authored ruleset could express. This high-dimensional feature space allows the model to distinguish subtle indicators of genuine fraud from the noise that had overwhelmed the rule-based system, providing immediately actionable insight on true versus false fraudulent activity.

The transition from rules to machine learning produced significant, measurable improvements across both dimensions of fraud management. The false positive rate fell by 50%, cutting in half the volume of legitimate transactions requiring manual review and freeing the fraud team for higher-value investigative work. Simultaneously, the fraud detection rate increased by approximately 60%, meaning substantially more genuine fraud is now intercepted before losses occur. Key outcomes:

• 60% increase in fraud detection rate — more genuine fraud caught at the point of authorization
• 50% reduction in false positives — direct reduction in investigation cost and analyst burden
• <300ms transaction scoring latency — no degradation to customer experience
• Fraud team now more effectively used; bank positioned for substantial operational savings

• Rule-based fraud systems have a fundamental detection ceiling — they only catch patterns analysts already know to look for. ML surfaces latent signals invisible to handcrafted logic.
• False positive rate is as consequential as detection rate; an unmanageable investigation backlog exhausts analyst capacity and can obscure genuine fraud signals.
• Sub-300ms real-time scoring enables point-of-authorization prevention rather than post-hoc detection, which is essential for card-present and mobile payment channels.
• Analyzing tens of thousands of latent features requires a platform purpose-built for financial transaction data — scale and feature richness are what unlock detection improvements that rules cannot achieve.