A running service can still be wrong
A demand forecast can continue returning numbers after customer behavior changes. Availability monitoring will not show whether those forecasts support useful decisions. Define quality indicators and the operational outcome they are meant to improve.
Watch the inputs and the workflow
Track missing fields, unexpected categories, and changes in data sources. Compare results across meaningful groups and time periods. When ground-truth labels arrive later, distinguish immediate health checks from delayed performance measurement.
Give someone responsibility
Specify who reviews alerts, who can pause the system, and how users report incorrect outputs. Record model, preprocessing, and dataset versions together. The paper Hidden Technical Debt in Machine Learning Systems is useful background on why the surrounding system deserves attention alongside the model.
Plan changes as experiments
Compare a candidate update with the current version on a fixed evaluation set and recent examples. Define a rollback path and a fallback workflow. Retraining is a possible response to a diagnosed problem, not a substitute for understanding whether the data source or business process changed.
Try this: Write an operating checklist with an owner, three quality signals, an escalation rule, and the steps for reverting a model release.