What is model maintenance?
Model maintenance is the work required to keep a model useful after it goes live. Model monitoring detects problems, and maintenance fixes them or prevents the predictable ones. The tasks include retraining on recent data when model drift erodes accuracy, adjusting confidence thresholds as volumes change, updating features when an upstream system renames a field or changes a code list, patching the libraries and serving software the model depends on, repeating model validation after significant changes, and retiring the model cleanly when it is replaced.
Retraining can run on a fixed schedule or be triggered by a monitoring alert. Each new version is recorded in a model registry and is usually tested against the current one in a champion-challenger comparison before it takes over.
Model maintenance in practice
A health system uses a model to identify discharged patients at high risk of readmission so care managers call them first. Its maintenance is planned a year ahead. Annual diagnosis code updates are mapped into the model’s features before they take effect, and an upgrade to the electronic health record triggers a full revalidation. The model is retrained twice a year and recalibrated whenever the share of patients flagged moves outside an agreed band.
Many AI projects are funded only through launch. Naming an owner and a maintenance budget at the start keeps the model from degrading quietly after the project team moves on.