Your AI scribe may be billing higher than you did. You still sign it.

What the AI scribe is actually doing to your codes

Ambient AI scribes listen to the visit and draft the note for you. That part is genuinely useful. The catch is in what a thorough draft does to your billing.

A 2026 JAMA Health Forum policy analysis and a brief in npj Digital Medicine both documented the same drift: notes produced with ambient AI tend to carry more documented elements than notes the same clinicians wrote by hand, and those extra elements push the visit toward a higher evaluation and management level. The case studies cited average claims running roughly 12 to 18 percent higher after adoption.

Read that carefully, because it is easy to misread as good news. A richer note is not the same as a sicker patient or a longer visit. The model is good at capturing review of systems, history detail, and assessment language. When it captures more of that, the coding logic reads a higher level, even if the underlying visit was the same fifteen minutes it always was.

Why a higher code is your problem, not the vendor's

Here is the part that matters for a licensed physician. The chart goes out under your name and your attestation. You are certifying that the documented work was done and was medically necessary. The scribe does not sign. You do.

So if the note documents a full history and high-complexity decision-making, and the actual encounter did not rise to that, the gap is now in your record with your signature on it. That is the exact shape of an upcoding finding. It does not require bad intent. A pattern of codes that run higher than your peers, unsupported by what the visit actually involved, is what triggers a payer audit and, in the worst case, a False Claims allegation.

The honest framing is this: the AI made the note look like more work. If you sign without checking, you have attested to work the visit may not support. The convenience moved the risk onto you.

It also helps to remember how audits actually find this. Reviewers rarely start with intent. They start with patterns: a clinician whose code distribution shifted upward after a clear date, with charts that read fuller but visits that look the same length. An ambient scribe rollout is exactly the kind of date that draws a line on that graph. You may have done nothing wrong on any single chart and still light up as an outlier across the panel, simply because the tool changed how your notes read. That is the moment your individual review habit becomes the thing standing between you and a records request.

The downward adjustment that is coming for you anyway

There is a second reason not to ride the higher codes. Payers and CMS are aware of the same data you are reading here, and the response under discussion is an automatic downward coding-intensity adjustment for practices that adopt scribes.

In plain terms: if scribe-using practices code higher on average, the payer can assume the lift is documentation-driven rather than acuity-driven and discount it across the board. You would not even see the individual decision. The reimbursement model simply shifts to expect inflation and correct for it.

That makes chasing the higher code a poor trade. You take on audit exposure today for revenue that the system is preparing to adjust away tomorrow. The stable position is to code accurately and be able to defend every chart, which is also the only position that survives an audit.

How to check the code before you sign

You do not need a coding degree to catch this. You need a short, repeatable check on every AI-drafted note, the same way our companion briefing on how doctors use AI for clinical notes safely treats the review step as non-negotiable.

• Read the elements the note claims. If it lists a full ten-system review of systems or a detailed past history, confirm those were actually asked and addressed, not auto-filled from the chart or inferred.
• Match the assessment to the visit. If the decision-making language reads more complex than the encounter felt, downgrade it to what you actually did.
• Sanity-check the level against your own gut. If the visit felt like a level three and the note is coding a level four, that gap is the thing to resolve before you sign, not after.
• Delete what did not happen. The scribe may document a normal exam finding you did not perform. If you did not do it, it does not belong in the note, and it certainly should not raise the code.

This adds well under a minute per chart. It is far cheaper than one audit.

A note on documentation while you do it. If you downgrade a level the scribe suggested, you do not need a paragraph explaining yourself. The corrected note is its own record. The point is that the chart, as signed, reflects the real visit, so that if anyone ever pulls it, the code and the documentation tell the same story. That alignment is the whole game. An upcoding problem is almost always a mismatch between what the note claims and what the visit supported, and the read-back is simply you closing that gap before it leaves your hands.

Where consent and disclosure fit

A quick boundary note. The billing-accuracy problem is separate from whether your patients know an AI is listening in the first place. If that question is on your mind, our briefing on when AI scribes need patient consent covers the disclosure side directly. This briefing is only about the code on the claim.

The skill under the tool

Every ambient scribe vendor will tell you the note is ready to sign. The quiet truth is that a note good enough to look ready is exactly the note most likely to carry a code you did not earn. The skill that protects you is not faster documentation. It is the trained habit of reading what the model wrote against what you actually did, every single time, before your name goes on it.

That habit is teachable and it transfers across every tool you will use. AI for Physician Notes builds the verify-before-you-sign workflow for exactly this, and the two minute course quiz will point you to the right starting place for your practice.