What post-acute teams get wrong about clinical AI

Every skilled nursing operator I talk to has run a pilot. Very few have run a second one. The postmortem usually blames the model — it missed cases, it cried wolf, the vendor oversold it. That explanation is comfortable and it is almost always wrong.

The pilot was in the wrong place, not wrong

A deterioration score that lives on its own dashboard is a score nobody sees. The nurse is in the eMAR. The DON is in the morning stand-up. The medical director is on the phone. If the output does not appear inside one of those existing moments, it does not exist.

We learned this the hard way building Qurvia. The first version had a beautiful risk panel. Adoption was near zero until we moved the signal into the documentation flow the nurse was already in and attached the note that produced it.

Traceability is the adoption mechanism

Clinicians do not distrust AI because it is AI. They distrust an assertion with no provenance, which is the same standard they apply to a colleague. Show the note, the vital, the med change that drove the signal, and the conversation shifts from "do I believe this" to "what do I do about this."

Somebody has to own the alert

An alert with no named owner degrades into wallpaper on a predictable schedule — roughly two weeks in our experience. Before a pilot starts, write down the role that acts on each output and what the action is. If you cannot fill that in, do not run the pilot.

What to measure instead

Count actions, not alerts. Orders placed, assessments completed, escalations made, transfers avoided. A system that fires half as often but changes care twice as much is the better system, and alert volume will tell you the opposite.