For Heads-of · Practitioner
Model Monitoring
Continuous observation of a deployed model's inputs, outputs, performance and behaviour, to detect degradation, drift, misuse or security anomalies.
- monitoring
- observability
- drift
- performance
Metrics
Accuracy or other performance measures against ground truth where it is available, input and output distributions, latency, error rates broken down by group, and anomaly flags.
Tools
Logging predictions, comparing them with baseline distributions, setting thresholds, tracking fairness metrics and alerting when a threshold is breached. AI system monitoring and logging is the control.
Triggers
Monitoring is only useful if a result leads somewhere: an investigation into why performance dropped, and a decision to retrain, roll back or escalate. Without that, drift and degradation are observed but not acted on.
Related Terms
Audit Trail
A complete, timestamped record of the actions an AI system took and the decisions made with it, so that they can be reviewed, explained and attributed afterwards.
Incident Response
A structured process for detecting, assessing, containing and resolving adverse events involving AI systems, such as model failure, a security breach or a fairness harm.
Model Drift
A change in the relationship between a model's inputs and the environment it was fitted to, such that the model's assumptions no longer hold. The model itself is unchanged; the world it was trained on has moved.
Performance Regression
A decline in model performance over time, typically caused by shifts in the data, concept drift or changes in the environment the model operates in.
Post-Market Monitoring
A provider's documented, continuous collection and analysis of data on how its high-risk AI system performs once in use, so that problems are found and fixed throughout the system's life. It is the legal form of model monitoring under the EU AI Act.