What it means
Model monitoring is the ongoing tracking of an AI system's performance, behaviour and usage after it goes live, to detect problems such as declining accuracy, changing input data, errors or misuse. It turns evaluation into a continuous process.
Real-world example
A city's AI system for classifying citizen reports (potholes, broken streetlights, litter) worked well at launch. Six months later, monitoring shows a rising share of misclassified reports, because a new reporting app changed how citizens describe problems. The team retrains the model and adds an alert for future drops in accuracy.
Why it matters for AI adoption
AI systems can degrade silently as data, users, providers' models or the world change. Monitoring needs defined metrics, responsible people and clear procedures for responding. When using external models, provider updates can also change behaviour without notice.
Questions to ask
- What do we monitor after launch, who reviews it, and how often?
- What happens when performance drops, and who decides whether to pause the system?