Why “Successful” AI Projects Die in Regulated Finance

AI pilots in regulated finance often fail not because models are wrong, but because workflows, evidence, and auditability are missing. This article introduces the REPLAY test (Record, Evidence, Policy, Lifecycle, Actions, Yield) and shows common failure modes in disputes and fraud: pending transactions, policy drift, false-positive harm, and GenAI narrative drift. You will get a scale-readiness checklist plus the operational metrics that predict whether a system will survive production.