AI Drafts. The Clinician Verifies. That Line Cannot Move.

A medication dose in a generated note read correctly. It was wrong. The patient had described a different number entirely, and the model had quietly rounded it to something more common, more plausible, and completely inconsistent with what she actually said.

I caught it because I read the whole note. That sentence should not be remarkable. In practice, it is the entire discipline this lesson is about.

Never Treat Generated Documentation as Automatically Correct

Speech recognition can miss words. Speakers can be confused with one another. Numbers can be transcribed incorrectly. A language model can infer something that sounds clinically reasonable but was never actually established in the conversation. None of these failure modes are rare or exotic. They are the ordinary cost of a system built on probability rather than certainty.

Where to Look First

Some parts of a note carry more consequence than others if they are wrong. Review these first, every time: patient name and identifiers, medication names and doses, allergies, gestational ages and other dates or measurements, diagnoses, pertinent negatives, procedures, follow-up instructions, and the assessment and plan itself.

Anywhere a wrong statement could change what happens to the patient next deserves a second look before it deserves a signature.

Proper Names Are the Weak Point

Modern systems can be surprisingly good with specialized medical terminology and medication names. Proper names remain the most fragile part of the pipeline. A practical habit is to spell unusual patient names or uncommon terms during the pre-encounter setup, when the workflow allows it, rather than hoping the system infers the spelling correctly from context.

Design Your Review Around Consequence, Not Convenience

Not every error deserves equal attention. A formatting quirk is annoying. A wrong allergy is dangerous. Your review process should be built around that distinction rather than around reading everything with the same level of urgency, which in practice usually means reading nothing with enough urgency.

Ask yourself, before you sign the next ambient-generated note, which specific line in that note would be the most expensive to get wrong. Read that line twice. Then read the rest of the note once.