Among a portfolio of eleven deployed models at a large system, the one with the highest sustained clinician engagement makes no prediction. It reads the chart and assembles a one-screen summary of what is already documented: active problems, recent changes, outstanding results, and what the last three notes said.

Engagement with the predictive models in the same portfolio ranges from moderate to negligible. The system's CMIO has a theory that is hard to argue with: the retrieval tool solves a problem clinicians experience continuously and can verify instantly. The predictive tools solve a problem they experience occasionally and cannot verify at all.

The tool required no clearance, since it makes no clinical claim, and it was built in-house in under four months on top of the system's existing data platform. It is also the hardest one to fund going forward, because it produces no measurable cost avoidance.

Several other systems have built something similar and describe the same pattern. It is a reasonable candidate for the most valuable and least celebrated category in clinical AI: software that finds the thing you already had.