Model monitoring is the ongoing practice of tracking how an AI system behaves after deployment. A model may look strong during testing and still degrade later because the data changes, the product changes, users behave differently, or the surrounding workflow shifts. Monitoring helps teams catch those problems early.
What Teams Monitor
Teams usually watch both technical and business signals. Technical signals can include latency, error rates, input distribution changes, output distribution changes, confidence shifts, drift, and tool failures. Business or workflow signals can include user satisfaction, conversion, escalation rate, moderation burden, or downstream error reports.
Monitoring matters because production is where AI meets reality. Real users are messier than benchmark datasets, and deployed systems interact with changing environments rather than fixed test sets.
Why Monitoring Is Different from Evaluation
Model Evaluation tells you how the system performed in a designed test environment. Monitoring tells you how it is behaving now. Both are necessary. A strong evaluation can reduce risk before launch, while monitoring catches the kinds of problems that only appear at scale or after conditions change.
Monitoring also supports incident response. If a model starts hallucinating more, missing key fraud cases, or making slower decisions under load, the team needs signals that reveal the problem before customers or regulators do.
Why It Matters for AI Literacy
Model monitoring is important because it helps readers understand that AI deployment is not a one-time handoff. A model is not finished when it goes live. It becomes part of an operational system that has to be observed, updated, and governed over time.
In that sense, monitoring is one of the clearest examples of AI as an ongoing responsibility rather than a static artifact.
Related concepts: Model Evaluation, Model Drift, Anomaly Detection, Model Card, and Responsible AI.