The Ambient AI Scribe Field Guide
You are the editor, not the writer. The note is not accurate because the tool is good. It is accurate because you reviewed it and signed it. The accuracy lives in the review step, not the software.
Open out loud. State patient, signalment, and presenting complaint at the start. "This is Bella, a 7-year-old spayed female Lab, here for vomiting."
Narrate the physical exam. If you did not say it, it is not in the transcript, and the model may fill the gap with a normal. Speak pertinent negatives too.
Slow down on numbers. Say drug names, doses, units, and frequencies clearly and once. Numbers are where transcription and generation errors converge.
Brief the owner. A client chatting to the pet all visit confuses speaker attribution. A light "I will be talking through my exam out loud for the record" helps.
Close with a 20-second recap. Before you stop recording, summarize the assessment, plan, medications, and follow-up in one pass. This anchors the sections the model builds from.
Read the physical exam section first. Fabricated findings hide there as plausible normals. If you did not do it or say it, delete it.
Verify every medication and number character by character. Drug names, doses, units, frequencies. Medication errors are the most common error category in published scribe evaluations.
Hunt omissions, not just fabrications. Studies find missing information is more frequent than invented information. Check the note against what you actually discussed and decided, not what reads as reasonable.
Confirm the assessment is your reasoning. Delete generic filler that sounds clinical but is not what you concluded.
Apply the default-to-wrong rule. Any finding, number, or plan item you do not remember generating is wrong until you confirm it.
Audit one random note weekly. Compare it against the transcript or your memory of the visit. Track the error patterns you find.
Bring specific failures to your vendor. "The PE template keeps inserting dental findings" gets fixed. "The notes feel off" does not.
Verify settings changes in the output. A toggled setting is not a fixed setting until you see it reflected in a real note.
Never sign on autopilot. Review discipline erodes as trust grows. The tool improving is not a reason to stop checking.