Fraud Model 99% Accuracy Paradox

Your Amazon payments fraud detection model shows 99.2% accuracy on your test set and passed all offline validation. However, after deployment, the fraud operations team complains it's 'completely useless' and wants to disable it. They report that 95% of flagged transactions are legitimate purchases, causing massive customer service volume. Meanwhile, confirmed fraud cases are still getting through. What's happening and how do you fix it?