AI system monitoring is the ongoing tracking of its behaviour after deployment into operation. It is necessary because output quality is not constant – it changes when the nature of the input data changes, when the provider updates the model, or when the environment in which the system operates itself shifts. Track three things. The quality of outputs on a regularly checked sample, assessed by a human. Operational indicators such as response time, error rate and consumption. And user feedback, including a simple way to report an incorrect output. Set a threshold at which the system pauses or switches to manual processing. Record findings; they form the basis for reassessing risks and for audit. Derive the frequency of checks from the system's impact: check more often for decisions affecting people than for supporting tasks.
See also: AI risk management, AI system incident, Human oversight of AI.
