A model that was accurate at launch drifts as the world changes. Data drift is a change in the inputs: a new customer segment, a renamed category, a sensor recalibrated. Concept drift is a change in the relationship between inputs and outcomes: the same signals start meaning something else. Both lower quality without raising a single error.
Monitor at three levels. Data checks catch broken inputs early: schema, missing values, ranges. Distribution checks compare live inputs and predictions with a reference window. Outcome metrics measure real quality, but labels often arrive late, days or weeks after the prediction, so the earlier checks are your warning system. Decide in advance what triggers retraining and who is paged.
Watch inputs, predictions and outcomes, and act on drift before labels arrive.