What clinicians correct most in AI scribe drafts is orders and meds, a 2026 study finds

The AI scribe draft and the note your provider signs aren't the same document, and a JAMIA study published May 12, 2026 counts where they split: orders, tests, and medications, the exact lines a CPT or HCPCS claim is built on. Researchers at the University of California, Irvine tagged every change clinicians made across 314 draft and final note-section pairs, and the edits pile up on clinical facts, not wording.

What the study counted

The method sets how much weight the numbers carry. This is a qualitative content analysis of paired drafts and signed notes, not a vendor accuracy benchmark and not an audit of coded claims. Nobody scored the drafts right or wrong. Eight researchers built a framework of 11 edit categories across three groups, clinical content, terminology, and language style, then tagged every edit in the sample of 200 encounters. They logged 713 tagged edits in all.

One detail governs how you read every percentage below. Because a single edit could fall in more than one category, the study reports each figure as the share of the 713 tagged edits carrying that label, and the authors say plainly the shares don't sum to 100%. So a 40.0% isn't two in five note sections. It's 285 of the 713 tagged edits.

What clinicians fixed most

The ranking is the finding. Orders, symptoms, medications, and diagnoses sit at the top. Wording, abbreviations, certainty, and tone sit at the bottom. The cosmetic edits are the rare ones. The ones that touch what a claim is built from are everywhere.

Why orders and meds are the coding exposure

Put the top of that table next to the coder's desk. Procedures, tests, and lab orders are the single most-edited content, and medications are close behind. Those are exactly the elements a professional or facility claim is built on, the CPT and HCPCS lines and the medication record. When a clinician swaps out what the draft said an encounter ordered, the draft narrative and the billable event have parted ways, and it's the draft narrative a downstream reader inherits.

That makes order and medication content the reconciliation you don't skip on an ambient-drafted encounter. A CPT or HCPCS code has to reflect the service actually performed and ordered, documented in the record, not a fluent sentence a model produced from the room audio. The check is the same one a careful biller already runs, now aimed at a draft: match the order and result content in the note against the order and result record itself, rather than trusting the narrative. In procedure-heavy specialties, where that content is densest, weight the review there first.

Diagnoses got edited too, but certainty barely moved

Diagnoses were edited in 25.9% of tagged edits, and revising the certainty of diagnostic and causal statements was one of five recurring edit themes the authors named. That side of the note matters for risk adjustment, and this desk has covered it. An over-confident draft that reads more certain than the visit was is a documented risk to how a diagnosis gets coded, and the mirror problem, a draft that quietly drops a chronic condition and costs a recapture, is the other half of it.

Keep it in proportion, though. Certainty and hedging edits were a small slice here, 29 of 713 tagged edits, 4.1%. That's consistent with the earlier finding that certainty drift is real but marginal in what clinicians touch. The bigger, less-discussed story in this study is the clinical-content edits above it, the orders, tests, and medications, which is what this piece is about.

Where a human coder is still required

The study stops at what clinicians edited, so treat what follows as our reading rather than its finding. The draft and the signed note both persist, so the gap between them is checkable before a claim goes out rather than after a payer asks. Matching the order and medication content against the actual order record is a review step, not a software feature. A tool that drafts from audio doesn't know which order was actually placed, and the clinician who corrected it was making a clinical call, not a coding one.

So the reviewer's job is the same as ever, just better aimed. The categories this study ranks tell you where an ambient draft drifts from the record, and that's where to spend the review. It's the discipline of confirming a condition was genuinely evaluated and addressed at the encounter, applied to a note a machine wrote first.

What this study does not establish

Be careful quoting these in a meeting. It's 200 encounters on 2 vendor systems at one health system, analyzed qualitatively, with moderate agreement among the reviewers who tagged the edits. The frequencies describe that sample. They aren't an industry error rate or a vendor comparison, and the paper never asks how many edits would have changed a billed code. What it does establish is narrower and more useful: clinicians revise specific categories of content at measurable rates, and the categories they revise most are the ones a code depends on.

That also puts a count under an argument this desk has made without one. The case that an AI-generated note still needs coder review before you keep what's on it used to rest on how the notes read. Now there's a tally of what the review step is catching.