There’s a number we watch more closely than almost anything else: how long it takes a practitioner to go from “note is drafted” to “note is signed off.” Not because it’s a nice metric for a slide deck, but because it’s the only honest test of whether ambient documentation is actually working.
Here’s why it matters more than accuracy alone. A note that’s 95% accurate but takes six minutes to review and correct isn’t a win over typing it yourself — it’s just a different way to spend six minutes. A note that’s 90% accurate but takes ninety seconds to scan and confirm is a real improvement, because the reviewing clinician is doing what they’re actually good at: catching the one thing that’s off, not reconstructing the whole visit from memory.
We aim for under three minutes for a standard visit, and most reviews come in well under that once a clinic’s templates are dialed in for their specialty. Getting there wasn’t about making the underlying model bigger. It was about three smaller things: structuring notes so the clinically important line isn’t buried in a paragraph, defaulting to the clinic’s own phrasing instead of generic clinical language, and being upfront when the model isn’t confident about something instead of quietly guessing.
That last one matters more than it sounds like it should. A note that states a vital sign with total confidence when the audio was genuinely unclear is worse than no note at all, because it looks trustworthy right up until it isn’t. We’d rather flag it and let the clinician fill in the one blank than have them re-verify every line out of habit.