The Review Pass That Makes an AI-Drafted Note Yours

The pitch delivered: the session ended, and a draft note was waiting. It reads clean. The objective section has numbers, the assessment sounds clinical, the plan is plausible. That polish is exactly the problem. When a tired clinician writes a weak note, the weakness is visible — missing data, a thin assessment, a sentence that trails off. When a language model writes a weak note, the weakness is fluent: a percentage that was never tallied, skilled reasoning flattened into activity description, yesterday’s assessment wearing today’s date. The moment you sign, none of it is the software’s draft anymore. It is your clinical record, your billing support, and your testimony if anyone ever asks. This article is about the review pass that happens between the draft and the signature — what specifically to hunt for, in what order, and how to keep the pass fast enough that you actually do it every time.

Medicare’s rule is old, short, and indifferent to how the note was produced: services must be authenticated by the author, by handwritten or electronic signature. And CMS has already answered the question of what happens when someone else drafts the note. In its 2017 Program Integrity Manual update on scribe services, CMS instructed reviewers to look only for the treating clinician’s signature — the scribe’s does not matter — because the clinician’s signature affirms that the note adequately documents the care provided. Swap the human scribe for a model and the logic does not move an inch. There is no lower standard for AI-assisted notes, no shared-blame arrangement with the vendor, and no reviewer anywhere who will weigh “the software wrote that part” as a defense. The draft belongs to the tool; the signed note belongs to you.

The professional associations have converged on the same position from the clinical side. APTA’s practice advisory on AI-enabled ambient scribe technology is direct: AI-generated clinical documentation may contain errors of omission or addition, or nuanced inaccuracies, and must always be reviewed. ASHA’s guidance on generative AI for clinicians frames these tools as support for clinical work, not a substitute for clinical reasoning, and points to the Code of Ethics obligation to evaluate any technology used in practice. AOTA’s journal has made the same call for critical evaluation of AI tools in occupational therapy practice. None of this is anti-AI — all three associations acknowledge the documentation burden these tools attack. It is a statement about where the review responsibility lands, and it lands on the signer.

A review pass that just re-reads the note will miss what matters, because AI drafts do not fail the way human drafts fail. A rushed human note has visible gaps. An AI draft has invisible substitutions — content that is grammatical, plausible, and wrong in ways that only the person who was in the room can detect. Three patterns account for most of them, and each one concentrates in a different section of the note.

1. Invented objective data

Language models complete patterns, and therapy notes are full of patterns that look like “producing target with 80% accuracy across 20 trials.” Run a session where you cued heavily but never tallied, and the draft may hand you a percentage anyway — not because the tool heard one, but because notes like yours usually contain one. The same mechanism invents trial counts, standardized score fragments, distances, and minutes. Transcription adds its own class of errors: left becomes right, the target sound migrates, a parent’s report in the waiting room becomes something you observed. The objective section deserves the most suspicion precisely because it is the section reviewers trust most, and because a fabricated number is the one error that reads as completely normal on a skim. The fix is a hard rule, not vigilance in general: every number in the draft is either data you collected today or it comes out. Deleting is not always enough — if no tally was taken, say what you observed and that formal data was not collected, which is an honest note rather than a hole.

2. Skilled rationale softened into bystander language

Medicare’s documentation standard for therapy turns on whether the record shows that the service required the skills of a therapist — that is the load-bearing wall in the Benefit Policy Manual’s therapy documentation requirements, and most payer review follows the same logic. A general-purpose summarizer does not know that wall exists. It describes what happened in the room the way a bystander would: “patient practiced word retrieval with support,” “completed transfer training with assistance.” Nothing in those sentences is false. They simply describe activity anyone could have supervised, which is exactly the reading a reviewer is looking for permission to make. Your session contained decisions — the cue level you chose and faded, the task you graded down and why, the compensation you disallowed because it would block the goal. The draft will reliably sand those decisions into “support” and “assistance.” The review question for the assessment and intervention lines is not “is this accurate?” but “could an untrained person have done what this sentence describes?” Wherever the answer is yes and the truth is no, rewrite the line to name the decision.

3. Copied-forward assessments

Tools that read your previous note for context inherit its gravity: yesterday’s assessment language drifts into today’s draft, sometimes verbatim, and a caseload documented by the same tool converges on the same sentences across different patients. This is the oldest documentation risk in electronic records wearing new machinery. Medicare’s review contractors have warned for years that entries worded exactly like or similar to previous entries are treated as cloned documentation that fails to support medical necessity, and HHS’s Office of Inspector General flagged copy-paste misuse as a fraud vulnerability back in 2013 — when it found that only about a quarter of hospitals even had a policy governing the feature. An AI drafting layer can reproduce that pattern at a scale no copy-paste habit ever reached. The tell is an assessment that would have been equally true last week. The defense is to ask, before signing: what does this assessment say that the previous note could not have said? If the answer is nothing, today’s skilled observation is missing — and a reviewer reading the chart in sequence will notice, because reading in sequence is what reviewers do.

Run the items in order, every time — the order goes where the risk is, numbers first. A fixed sequence is what makes the pass fast: you stop re-reading the note as prose and start hunting for specific defects, the way you scan an EOB or a schedule. On a routine session note, most items take seconds. The pass earns its keep on the items that occasionally take longer, because those are the catches that matter.

A review pass that waits until nine at night inherits the exact problem AI drafting was supposed to solve — and it reviews against a memory that has already merged three sessions together. The checklist only works when it runs close to the session it checks. The second half of the discipline is feedback: the same tool making the same error in every draft is not a review problem, it is a configuration problem, and the review pass should be shrinking over time as you fix the source.

1. 01

   Review at the point of memory

   Run the pass before the next session starts, or in the day’s documentation block at the latest. The accuracy items — invented numbers, missed deviations, borrowed observations — are only checkable while you still remember the session as itself, not as part of a blurred afternoon.

2. 02

   Run the checklist in its fixed order

   Same sequence every note: numbers, time, attribution, transcription details, skilled language, assessment freshness, the diff, omissions, plan, auditor read. Order builds speed; improvisation rebuilds the slow re-reading habit the checklist exists to replace.

3. 03

   Keep a two-week catch log

   Every edit you make before signing, tally it by failure mode. Two weeks of tallies tells you what your particular tool invents, softens, and copies — which turns a generic review pass into a targeted one, and gives you the evidence for the next step.

4. 04

   Fix the tool, not just the note

   Recurring catches go upstream: adjust the tool’s template, instructions, or settings so the error stops being generated — many tools accept custom guidance like “never state accuracy percentages unless dictated” or “flag data not explicitly stated.” If the vendor offers no way to fix a recurring failure, that is product feedback worth sending, and worth remembering at renewal.