How would you detect behavior drift in a production AI feature?

I maintain control charts or comparable baselines for important slices and validate the measuring system with periodic human labels. Provider aliases are resolved to observable model versions where possible, and unannounced external change is handled as a dependency risk. Response options include routing to a known model, reverting prompts, disabling a source or tool, narrowing supported behavior, or increasing review. After diagnosis, new representative cases and monitoring rules capture the failure without overfitting to one incident.