The hospital AI committee is now a real job. Most of them still meet like a book club.

Almost every large hospital system now has something called an AI governance committee. Very few of them can say, without checking, how many models are running in their own clinical environment this morning.

That gap is the story. The committees were created quickly, often after a board asked an uncomfortable question, and most inherited the shape of a research ethics board: periodic meetings, written submissions, a decision at the end. Clinical software does not behave that way. It drifts, it gets updated by the vendor, and it fails quietly.

What the effective ones borrowed

The systems that have made this work borrowed from medication safety rather than from research ethics. That means a register of everything in use, a named clinical owner per model, defined monitoring metrics, and a standing route to switch something off without convening anyone.

It also means accepting that the vendor's validation study is evidence about somebody else's patients. Local performance checks on local data are the only thing that answers the question a clinician actually asks, which is whether this tool is right about the people in front of them.

Where it still breaks

Three failure modes recur. Models that arrive inside an electronic record upgrade and never pass through governance at all. Monitoring that is defined but never staffed. And ownership that sits with informatics rather than with the clinical service that carries the consequence.

None of these are technology problems. They are the ordinary operational problems of adding anything new to a hospital, and they respond to the ordinary operational answers: a register, an owner, a metric, a review date.

What to ask your own institution

How many models are live. Who owns each one. What number would trigger a review. Who can suspend it before lunchtime. An institution that can answer all four is further ahead than most of the published frameworks would suggest.