COMPEL Glossary / model-monitoring
Model Monitoring
Model monitoring is the continuous, automated observation of AI models operating in production to track performance metrics (accuracy, latency, throughput), detect data drift and concept drift, identify anomalous behavior, monitor fairness metrics, and ensure the model continues to operate within the acceptable parameters defined by its governance framework.
What this means in practice
Without monitoring, models can silently degrade over weeks or months as the real world changes, producing increasingly inaccurate or biased results while appearing to function normally. For organizations, model monitoring is the operational foundation of responsible AI because governance policies are meaningless if violations go undetected. In COMPEL, model monitoring is assessed under the Technology pillar during Calibrate and designed as a core component of the operational infrastructure during Module 3.3, with agentic-specific monitoring requirements covered in Module 2.5.
Why it matters
Governance policies are meaningless if violations go undetected. Model monitoring provides the operational foundation of responsible AI by continuously tracking performance, detecting drift, identifying anomalous behavior, and monitoring fairness metrics. Without monitoring, organizations cannot verify that their AI systems continue to meet the standards established during initial approval, creating growing but invisible governance gaps over time.
How COMPEL uses it
Model monitoring is assessed under the Technology pillar during Calibrate and designed as a core component of operational infrastructure during Module 3.3. During Model, monitoring requirements are specified for each AI system. The Produce stage implements monitoring dashboards and alerting. The Evaluate stage assesses monitoring effectiveness by reviewing whether degradation was detected promptly and governance violations triggered appropriate responses.
Related Terms
Other glossary terms mentioned in this entry's definition and context.