Five signals. No vanity metrics.
Health is scope confidence, quality evidence, real adoption, cost control and assurance standing — scored equally, because an AI programme fails the moment any one of them is ignored.
Adoption healthy in two regions; third region blocked on support training.
Retrieval quality plateaued — golden set refresh scheduled.
Benchmark programme. Reuse its eval harness across the portfolio.
Cost per task 7x target. Needs caching and model routing before scale.
Legal sign-off on data provenance is the critical path.
Amber is information, not failure
A pilot should be amber. Treat red on assurance or cost as the urgent ones — they block scaling hardest.
Adoption lags quality by a quarter
Strong quality with weak adoption is an enablement problem. Do not respond by retraining the model.
Cost per task is a leading indicator
If it is not falling as volume rises, your architecture is not ready to scale regardless of how good the output is.