Before you sign it.
A well-written wrong note is more dangerous than an obviously bad one, because it does not trigger the doubt that makes you read carefully. This is the review discipline, in order, built around how ambient documentation actually fails.
01 Why fluency is the problem
A note written by an exhausted resident at the end of a shift reads like one. The prose is clipped, the structure wobbles, and your brain flags it for attention automatically. You read it properly because it looks like it needs reading.
An ambient AI note reads like a careful clinician wrote it with time to spare. Clean structure, complete sentences, correct terminology, appropriate register. And that fluency suppresses exactly the scrutiny reflex you need.
You cannot rely on a note feeling wrong, because it will not. Ambient documentation moves review from an instinct to a procedure. If your check is "read it through and see if anything jumps out", you will miss the errors that matter, because none of them jump out.
02 The errors that survive a read-through
These are not random. They cluster in predictable places, which is what makes a checklist possible.
| Error type | What it looks like in the note |
|---|---|
| Dropped negative | You asked about chest pain and the patient denied it. The note simply does not mention chest pain. Absence is invisible on a read-through. |
| Certainty inflation | "Probably viral, review if worsening" becomes "viral illness". The hedge was the clinical content. |
| Speaker attribution | In a consultation with a family member present, something the relative said is recorded as the patient's report. |
| Discussed-as-confirmed | A differential you raised out loud and dismissed appears in the assessment as though it stands. |
| Temporal collapse | "Three days" and "three weeks" are one syllable apart in speech. Onset timing is a high-frequency error. |
| Laterality and dose | Left/right, and any number attached to a medication. Small tokens, large consequences. |
| Plausible completion | A gap in what was heard, filled with the clinically typical thing. The most dangerous category, because it is indistinguishable from correct content by style. |
Which is more likely in an ambient note: a symptom the patient never mentioned appearing in the history, or a symptom the patient did mention being absent from it?
Choose, then open.
Omission is considerably more common — and considerably harder to catch. Reviewing a note, you are reading what is on the page. Nothing on the page prompts you to notice that the patient's report of intermittent dizziness is not represented anywhere.
This is why the review order below starts with recalling the encounter before reading the note. Once you have read a fluent, complete-looking document, your memory of the consultation quietly reshapes itself to match it. Reading first contaminates the only check you have for omission.
03 The review order
Run these in sequence
- Before reading: recall the encounter for five seconds. What were the two or three clinically important things? Now read and confirm each appears.
- Check the negatives. What did you rule out by asking? Confirm each pertinent negative is present.
- Check every number and side. Doses, durations, onset timing, laterality, vital signs. Verify against your own recollection or the source.
- Check the hedges. Anywhere you were uncertain, confirm the note is uncertain too. Certainty inflation is the most consequential silent error.
- Check attribution if anyone else was in the room.
- Read the assessment and plan last, and slowly. This is the part that drives care and the part read by everyone downstream.
- Ask what is missing — one deliberate beat before signing.
04 Report the errors
When the system gets something wrong, tell whoever owns it. This is the step almost everybody skips, and it matters more than it looks.
Ambient systems are tuned. Patterns of error — a specialty term consistently misheard, a note structure that keeps dropping a section — get fixed when they are reported and persist indefinitely when they are not. A clinician who silently corrects the same error two hundred times has done two hundred units of work and produced no improvement for anyone, including themselves.
Find out who receives that feedback where you work. Most clinicians cannot name them, which tells you how rarely the loop is closed.
05 The honest trade
Ambient documentation converts composition work into verification work. That is a good trade — verification is faster and less draining than drafting from a blank page.
It is only a good trade if you actually do the verification. A clinician who signs unread has not saved time; they have transferred risk onto a patient and onto their own signature. The tool's entire safety case rests on a review step that only exists if you perform it.
This week's challenge
For five consecutive notes, write down — before opening the draft — the three most clinically important things from the encounter. Then read the note and check all three appear correctly. Most clinicians find at least one omission in five, which is both reassuring about the tool and clarifying about why the review exists.