Ambient AI scribes omit more than they invent, and an omitted chronic condition is a lost HCC

Most ambient AI scribe notes are clean. The residual error that isn't clean tends to be the note leaving something out, not the note making something up, and that is the error a coder is positioned to catch. Two peer-reviewed evaluations published in 2026 landed on that same pattern from opposite directions.

The first is a pragmatic pilot in JMIR Medical Informatics (published April 17, 2026), in which 31 physicians used an ambient AI scribe to draft 7,545 clinic notes and reviewed a sample of them for four error types. The second is a vendor-neutral comparison in the Annals of Internal Medicine (June 2026), which scored AI-generated notes against human-written notes across five standardized primary-care cases. Neither study is about diagnosis codes. The risk-adjustment reading below is ours, not theirs.

What the two studies actually found

The JMIR pilot ran from July to August 2024. Physicians evaluated 356 (4.7%) of the AI-generated notes and rated each error's potential to harm a patient if left uncorrected. The headline is reassuring: the authors report that 94.7% of reviewed notes were free of significant errors.

The part that matters for a coder is the breakdown of the errors that did occur. Accidental omissions were the most frequent error type, in 18% of reviewed notes. Fabrication was rarer: hallucinations showed up in 11.5% of reviewed notes, and accidental inclusions, which carry real but misplaced content rather than invented content, in 9.3%; bias was rare at 1.1%. The authors' own conclusion is that "careful clinician review of notes remains imperative."

The Annals evaluation, funded by the Veterans Health Administration, took a different route to a similar place. Eleven AI scribe tools, 18 human note-takers, and 30 blinded raters worked from five audio-recorded standardized cases. Raters scored every note on the modified Physician Documentation Quality Instrument (PDQI-9), a 10-domain, 50-point scale. The authors report that human notes scored higher than AI notes across all five cases, with the largest single-case gap in acute low back pain (human 43.8 vs. AI 20.3). In the pooled domain analysis, AI scored lower on all 10 domains, and the three largest deficits were being thorough, organized, and useful.

Read together, one study says the most common ambient-AI error is omission, and the other says the biggest quality gap is thoroughness. Those are the same finding in two vocabularies.

Why omission is the coder's problem and not the physician's

A physician editing an AI draft is comparing it against the visit they just had. They know the patient's COPD came up, so if the draft skips it they can add it back. That is the correction loop both studies assume, and it mostly works.

A coder never had the visit. A coder has the note. If a condition was discussed and the note left it out, there is nothing in the record to react to, no flag, no discrepancy, no cue. A hallucinated diagnosis is at least visible: a coder checks it for support under the MEAT criteria, finds none, and drops it. An omitted diagnosis is invisible by definition. It is the one error class the downstream coder cannot catch by reading harder, because the thing to catch isn't on the page.

That inverts the usual worry. The compliance conversation about AI notes has centered on the tool adding an unsupported code, which is the failure mode behind the ambient-scribe risk-adjustment exposure and the capture-delta audit. Omission is the mirror image. For risk adjustment it costs money, and it does so silently.

For risk adjustment, a dropped chronic condition is a lost recapture

Under the CMS-HCC V28 model that pays in 2026, most chronic conditions have to be documented and coded every calendar year to keep contributing to the patient's risk score. A condition that was real, was discussed, and simply never made it into the note is not an audit problem. It is a condition that goes uncoded, so its HCC and its RAF weight fall off the risk score for the year. That is the recapture gap, and it is the opposite of the overcoding story: nobody gets a letter, the money just isn't there.

The conditions most exposed are exactly the ones a clinician mentions in passing on a stable patient: the long-standing diabetes that isn't today's chief complaint, the CKD noted from a lab trend, the COPD that is controlled and barely comes up. Those are the lines an ambient draft is likeliest to compress out, and each one carries an HCC.

The chronic conditions most likely to get dropped, and what they carry

Every code and HCC below was verified against the CMS-HCC V28 model for payment year 2026. Community weights are shown; the weight your plan realizes depends on the enrollee's segment.

None of these is an edge case. They are the everyday chronic load of a Medicare Advantage panel, and they are the conditions a recapture pass exists to catch. An ambient note that drops one is a recapture you have to find another way.

The completeness cross-check, since the studies stop at naming the problem

The QA habit for AI notes has mostly been "read the note against the visit." A coder can't do that. So build a check the coder can actually run, and run it against the two things a coder does have: the patient's own history and the note in front of them.

• Reconcile against last year's captured HCCs. Pull the conditions that mapped to an HCC for this patient in the prior year. For each one, confirm it is either addressed in this year's note or has a documented reason it isn't (resolved, ruled out, no longer active). A chronic condition that was captured last year and is simply absent this year, with no explanation, is the omission to chase down before you move on.
• Reconcile against the problem list. Where the active problem list carries a chronic condition and the AI-drafted note is silent on it, that is a query, not a code. Do not code from the problem list alone, and do not treat the note's silence as a resolution.
• Send it back to the provider. The fix for a suspected omission is a provider query, because the coder can flag that a captured chronic condition went unaddressed but cannot document that it was addressed. Reconciling is the coder's job; re-documenting is the provider's.

This is the same discipline as the QA pass every AI-drafted note still needs, narrowed to the one failure mode these two studies say is the most common: the condition that isn't there.

What the studies did not test

Keep the limits straight, because neither study measured coding at all. The JMIR work is a single-health-system pilot in which physicians self-evaluated a 4.7% sample of their own tool's notes, so the omission rate is an estimate from one setting, not a benchmark. The Annals evaluation used simulated standardized cases, and the authors note the human notes were not produced under real-world time pressure, which flatters the human comparison. Neither study reported an ICD-10-CM code, an HCC, or a RAF figure. The connection from "AI notes omit content" to "an HCC goes uncaptured" is a coding inference from the guidelines and the V28 model, not a measured result, and it is worth stating plainly to anyone who asks.

Where a human coder is still required

An omission has no signature. There is no confidence score on a note that flags the sentence that should have been there, and no fluency metric that penalizes the tool for being too brief about a stable chronic condition. The only thing that surfaces a dropped diagnosis is a person reconciling the note against what the patient is already known to carry.

That reconciliation is also the work that decides whether the risk score is complete and whether the documentation behind each condition will hold when someone asks why the code is, or isn't, on the claim. A tool that quietly leaves out the COPD hasn't made an error the tool can see. It has made one only a coder can see, and only if the coder goes looking.