What to check first when you review an AI-drafted note

When I sign a note, I'm saying it is an accurate record of the visit. That holds whether I typed the note or software drafted it. CMS says so in its guidance on signatures: if you use a scribe, including artificial intelligence technology, you sign the entry to authenticate the documentation and the care you provided or ordered. You don't need to record who or what transcribed it.

So the review is the job. The question is how to do it well in the minute or two between patients, and my answer is to stop reading the draft from top to bottom.

Where the errors are

Errors in AI drafts aren't spread evenly through the note. In a 2025 study in npj Digital Medicine, Asgari and colleagues had doctors label 12,999 sentences from notes that GPT-4 drafted from mock primary care consultations. They found a hallucination in 1.47% of sentences and an omission rate of 3.45%. The reviewers rated 44% of the hallucinations as major, meaning they could affect diagnosis or management if no one corrected them. Major hallucinations turned up most often in the plan (21%) and the assessment (10.5%).

By type, 43% of the hallucinations were fabrications, 30% were negations and 17% were contextual errors. The other 10% were causal claims the conversation didn't support. Most major omissions (55%) concerned the patient's current issues.

That study used a different model and simulated visits, so its rates don't transfer to any one product, ours included. What transfers is the map. Read the plan and the assessment most closely. Then look for negatives that have flipped and for problems the patient raised that never made it into the note.

The order I use

1. Plan, line by line. Check each drug with its dose, route and frequency, then each test, referral and follow-up interval, against what you said in the room. An error here changes what happens to the patient next.

2. Assessment. Is each diagnosis one you actually made? Check laterality, and look for causes you never stated, such as "likely due to" a medication you didn't blame.

3. Negatives. Find every "denies," "no" and "without" in the history and confirm each one. A flipped negative reads exactly like a correct sentence until you compare it with what was said.

4. Numbers. Vitals, scores, durations and dates. Tap any line you doubt to jump to that moment in the transcript.

5. Missing problems. Go back to the reason for the visit and the patient's list of concerns, including the one raised with a hand on the door. Each should appear somewhere in the note.

6. What you didn't say out loud. Exam findings and observations you made silently won't be in the draft. Charted flags those sections so you can fill them in before you sign.

Lines Charted couldn't hear clearly are highlighted, so confirm them as you reach them. Tone and phrasing come last, because they rarely change care.

Making it faster

Two habits stood out among the clinicians in our pilot who signed fastest. They said the plan out loud before the patient left: "We're going up to 20 milligrams of lisinopril, and I'd like a blood test in two weeks to check your kidneys." That doubles as teach-back for the patient and gives the draft a clear source. They also signed each note before opening the next chart, while the visit was still fresh.

Two months into the pilot, most edits were in the plan, and they were usually details the clinician had decided but hadn't said aloud.

When the same error keeps coming back

If a drug name is misheard at every visit, or a finding keeps landing in the wrong section, tell your implementation lead instead of correcting it by hand each time. Repeated errors usually point to something in the template or setup that we can change for the whole clinic.