Financial fraud detection is commonly evaluated as a stationary classification problem even though transaction behavior, fraud tactics, reporting practices, and defensive controls change over time. This study examines the consequences of concept drift for predictive performance, drift detection, explanation stability, and governance. A temporally ordered financial transaction stream of 9,000 observations was constructed from a documented financial simulation design, with 14 behavioral and transactional predictors and 530 fraud events. The experiment imposed identifiable sudden, gradual, incremental, and recurring changes after an initial training and validation period. Three static models, logistic regression, random forest, and histogram gradient boosting, were compared with online logistic regression, sliding-window random forest, and drift-triggered random forest under prequential evaluation. Page-Hinkley, DDM, KSWIN, and an ADWIN-style window detector were assessed using detection delay, false alarms, and missed changes. Explanation stability was measured through permutation importance rank agreement across pre-drift, post-drift, and late-stream periods. Static random forest achieved mean PR-AUC of 0.192, while online logistic regression achieved 0.109 but substantially higher mean recall, 0.596 versus 0.234. None of the adaptive models produced a statistically significant PR-AUC improvement over static random forest. Page-Hinkley generated fewer false alarms than the more sensitive window detector, while DDM and KSWIN missed all prespecified changes under the selected operating conditions. Explanation rankings were unstable after sudden drift. The findings show that adaptation is not automatically superior. Effective fraud operations require selective retraining, threshold governance, explanation review, analyst validation, and explicit decision rights. The paper develops the Adaptive Fraud AI Governance Framework to connect technical monitoring with institutional accountability.