Before AI Hits the Ward, Doctors Test It

Sonographers do the same wrist motion, scan after scan, all shift. It's a known cause of repetitive strain injury. Nobody built them an AI tool. Not yet. But UnityPoint Health is now asking whether one should exist for them too, alongside the more familiar targets: the note, the chart, the alert.

That question, and who gets to answer it, is the interesting part.

The 60-physician filter

UnityPoint Health has put together a standing group of 60 physicians, spread across specialties and markets, whose job is to test AI tools before they roll out system-wide. Not after. Before. The idea is simple: a tool that looks good in a vendor demo can still fail on a Tuesday afternoon with a full list and a patient who won't stop talking. The only way to know is to hand it to a doctor who has that Tuesday afternoon and ask what broke.

Most health systems don't do this. They buy the tool, mandate it, and measure adoption. UnityPoint is measuring something upstream of adoption: does this survive contact with an actual shift.

What's on the table

The tools in front of that 60-physician group aren't exotic. They're the usual list, which is exactly the point:

• Ambient scribing, so the physician isn't typing while the patient is talking.
• Chart summarization, piloted with Evidently, meant to surface what actually matters in a long chart instead of making the doctor scroll for it.
• Real-time capacity tracking across operating rooms, infusion spaces, and procedural areas, so staff aren't guessing which room is free.
• Alert design that subtracts noise rather than adding another popup to click through.

Gregory Johnson, UnityPoint's chief medical officer, put the goal in one line: doctors are no longer "attached to the keyboard." That's a low bar to state and a hard one to clear. Every EHR vendor has promised some version of it for a decade. What's different here isn't the promise, it's the mechanism for checking whether it's true: a physician panel with the standing to say a tool isn't ready.

Why this is the harder problem

Most clinical AI coverage is about capability: can the model read the scan, draft the note, catch the deteriorating patient. Capability is necessary and increasingly available. What's scarce is the second question: does this tool earn a place inside a real shift, where the doctor is constantly interrupted and has seconds, not minutes, to decide whether to trust what's on the screen.

That second question can't be answered by a benchmark. It can only be answered by watching a doctor use the thing, on a bad day, with a full ward, and seeing whether it saves time or adds a new kind of vigilance tax. UnityPoint's 60-doctor group is, in effect, a verification layer between what a vendor claims and what a hospital deploys.

That's the same bet Medows is built on: an AI workspace only earns trust if a doctor can check its work in the moment, not just admire its accuracy in a press release. A tool that's usually right but opaque about when it's wrong is not a tool a doctor can safely rely on mid-shift. The fix isn't a better model. It's a workflow that lets the doctor verify, fast, without breaking their round.

Ambient scribes and chart summarizers will keep shipping. The systems that get real traction won't be the ones with the best demo. They'll be the ones that survived a physician panel that was allowed to say no.

Sources

• 6 Ways UnityPoint Health Is Trying to Make Clinicians' Jobs Easier (MedCity News, Aug 5, 2026)