Model Drift
Also called: Drift, Data Drift, Concept Drift
Model drift is the gradual decay in a model's performance as the world it operates in moves away from the data it was trained on. Nothing inside the model changes. Customer behaviour, pricing, product mix or fraud tactics change around it, and predictions that were accurate at launch become steadily less so. The decline is quiet, because a drifting model keeps producing confident outputs in exactly the right format.
In practice
Drift is the argument against treating an AI deployment as a capital project that finishes. Budget for monitoring, a retraining schedule and a named owner who watches the accuracy figure, and in diligence ask a target when each production model was last retrained and against what measure.