AI Scribe and E/M Coding: Does It Help You Code More Accurately?

The coding question vendors don't answer cleanly

When AI scribe vendors talk about ROI, the time savings are the headline number. The coding implications usually get a soft mention: "thorough documentation supports appropriate billing." That's true but it leaves the actual question unanswered.

Does the AI scribe help you code more accurately? Does it nudge you toward higher codes that aren't justified? Does it improve your audit defensibility? Does it change anything at all about how you should bill?

This article answers those questions specifically. What the AI's documentation actually does for E/M coding under the 2021/2023 guidelines, where the real gains are, and where the risks hide.

The 2021/2023 E/M coding framework, briefly

In the US, the 2021 Office E/M guidelines (extended in 2023 to hospital and other settings) restructured how levels are chosen. Code selection now is based on either:

1. Medical Decision Making (MDM): a structured assessment of the number and complexity of problems addressed, data reviewed, and risk of complications.
2. Total time on the date of service: face-to-face and non-face-to-face time spent on the patient's care.

History and exam still need to be documented, but they no longer drive code selection. The shift was supposed to reduce documentation burden, and it has, modestly.

AI scribes intersect with this framework in specific places:

• MDM determination: the AI's note structure affects whether the documentation supports the MDM level you chose
• Time documentation: the AI captures visit duration but may not naturally surface it in the chart for time-based billing
• Risk documentation: the AI's plan section often describes risk factors (medication considerations, monitoring needs, social determinants); how it documents them affects MDM scoring

Where AI scribes actually help coding

Three places where AI-generated notes tend to support better coding than typed notes:

The plan section is usually more thorough. Providers typing notes often abbreviate the plan ("continue current regimen, follow up 3 months"). AI scribes capture the conversation more completely, including the reasoning the provider expressed verbally ("we'll continue current regimen because labs are stable, follow up in 3 months unless symptoms recur, will increase if A1c rises"). The fuller reasoning supports MDM scoring more clearly.

Problem complexity is captured more accurately. Providers naturally describe complexity when they're talking to the patient ("this is more complicated than it looks because the symptoms could be from any of several causes") but rarely transcribe that reasoning fully into a typed note. AI scribes capture the reasoning, which supports the "moderate complexity" or "high complexity" MDM determination.

Counseling time is documented more thoroughly. Time-based billing requires documentation of the time spent counseling or coordinating care. AI scribes capture the verbal counseling, which both supports the time-based code and demonstrates the substance of the counseling.

A typical pattern: a practice that was billing 99213s for visits that could have supported 99214s with better documentation finds that AI scribe notes support the higher code legitimately. The bill goes up; the documentation is improved; the audit defensibility is improved.

Where AI scribes don't help (or hurt)

Three places where the relationship is more complicated:

Over-thorough notes that don't match actual MDM complexity. The AI scribe's note for a simple ear infection visit might be five paragraphs long with detailed assessment. The documentation looks like it supports a 99214 even though the MDM was 99213-level. A provider who signs the AI's thorough note for a 99213 visit may be tempted to code the higher level because the documentation "supports it." This is upcoding, even if the documentation looks adequate, because the MDM was actually lower than the documentation implies.

Phantom data review. AI scribes sometimes generate text suggesting the provider reviewed labs or imaging when the provider didn't explicitly mention doing so. If the note says "labs reviewed, stable" but the conversation didn't include a labs review, the documentation is incorrect. Auditors increasingly look for these phantom data reviews because they're a common AI scribe artifact.

Inconsistent time documentation. Some AI scribes record visit start and end times accurately; some don't. If the visit was 20 minutes but the documented time was 30 minutes (because the recording started before the patient sat down), time-based billing has a gap between documentation and reality.

The pattern that's risky: assuming the AI's thorough-looking documentation supports a higher code automatically. The provider's MDM still has to support the code; the documentation supports that MDM doesn't generate it.

What auditors look for in 2026

Major payers (CMS, large commercial insurers, MAC-level auditors) have started developing audit criteria that account for AI-generated documentation. The patterns they're looking for:

Notes that are uniform across providers. Every provider in the same specialty starts to write almost identical-sounding notes because they're using the same AI platform with similar defaults. Auditors who notice this pattern can look for evidence that the documentation actually reflects different patients' conditions versus boilerplate.

Notes that don't match the visit length. A 10-minute visit with a 1,200-word note generated by the AI looks suspicious. The auditor's question: did the provider actually spend 10 minutes on the content the note describes? AI-generated thoroughness without time-equivalent visit length is a coding red flag.

Notes with internally inconsistent content. "The patient denied chest pain" appears in the HPI, but "discussed chest pain workup" appears in the plan. AI scribes occasionally generate notes with this kind of internal inconsistency; auditors look for it.

Notes with high billing levels but thin MDM elements. A 99214 requires documented moderate MDM. The AI may produce notes that look thorough but actually have low MDM (one stable chronic condition, no significant data review, low risk). Auditors check the MDM elements, not just the note length.

The protection: provider review that ensures the note matches the visit. The AI generates a draft; the provider's review is what makes the documentation accurate and audit-defensible.

The vendor's role in coding accuracy

Some AI scribe vendors have started offering coding-specific features:

Automatic MDM scoring. Some platforms analyze the generated note and suggest an MDM level. This can be useful as a sanity check; it can also be misleading if it suggests higher levels than the actual visit complexity supports.

Time documentation prominently displayed. Some platforms surface visit duration alongside the note, supporting time-based billing.

E/M code suggestions. A few platforms suggest specific E/M codes based on the note. This is the most risky feature; the platform doesn't see the visit, it sees the note, and a thorough-looking note may suggest a higher code than the MDM actually supports.

The honest evaluation: code suggestion features from vendors are a starting point, not a billing decision. The provider's clinical judgment still drives code selection. A vendor that markets the coding feature aggressively is one to evaluate carefully; the marketing may be ahead of the responsible use.

What practices are actually seeing on billing

Across practices that have adopted AI scribes, the billing impact patterns:

Modest legitimate uplift for under-coders. Providers who were previously under-documenting and under-coding (typically the most rushed providers) often see a modest, defensible uplift. A provider who was billing 99213s for visits that supported 99214s starts billing the 99214s correctly.

No change for accurately-coding providers. Providers who were already documenting and coding accurately don't see a billing change. The AI saves them time without changing the bill.

Risky over-coding for misusers. Providers who treat the AI's thorough-looking note as license to bill higher codes regardless of actual MDM see short-term billing gains and long-term audit exposure.

The pattern that holds: AI scribes don't make you a better coder; they make you a faster documenter. Coding accuracy depends on the provider's discipline about matching the code to the actual visit, not to the documentation that happens to look impressive.

The audit defensibility upside

Setting aside the risks, the defensibility upside is real. A provider in a payer audit with AI-generated, thorough notes that match the visit is in a stronger defense position than a provider with abbreviated, typed notes that lack detail.

The reasons:

• More content for the auditor to evaluate
• Verbal reasoning captured (the provider's thinking, expressed during the visit, is on paper)
• More complete plan documentation supports more accurate retrospective interpretation
• Time documentation, if present, supports time-based billing claims

The condition: the documentation has to reflect the actual visit. An AI-generated note that diverges from what actually happened is a defensibility liability, not an asset.

Specific code-level considerations

A few specific codes where AI scribes affect billing patterns:

99214 vs. 99213. The most common borderline. AI scribe documentation often legitimately supports 99214 for visits that providers were previously coding as 99213. The lift is defensible if the MDM actually supports it.

99215. High-complexity visits. AI scribe documentation can support 99215 well, but the MDM has to be there. Multiple problems, significant data review, high risk.

99417 (prolonged services). Time-based code for visits over the threshold. AI scribes that capture total time accurately make this easier to bill correctly. Without the time documentation, providers under-bill these visits.

Counseling-dominant visits. Visits where more than 50% of the time was counseling can be coded by time alone. AI scribes that capture the counseling content support these codes well.

Annual wellness visits (G0438, G0439). These have specific required elements. AI scribe templates can include them; some platforms have AWV-specific templates that ensure all required elements are present.

The recommendation

For most practices, the right posture toward AI scribes and coding is:

1. Don't change your coding behavior just because the documentation looks more thorough. Code the MDM, not the documentation. If the MDM is 99213, code 99213.

2. Do allow yourself to legitimately uplift when the documentation accurately reflects higher MDM. Don't reflexively stay at low codes if the AI's documentation captures the actual complexity better than your typing did.

3. Be specific about time documentation. If you bill time-based codes, configure the AI scribe to surface visit duration prominently. Verify the duration matches reality.

4. Build provider review into the workflow. The review is what protects you from upcoding errors and from documenting phantom content. The review is the audit defense.

5. Don't market AI scribe adoption as a billing increase initiative internally. Providers under pressure to bill higher will misuse the documentation. The internal framing is "more thorough documentation, more time back, billing reflects the actual care." Anything more aggressive invites trouble.

Done this way, AI scribes are a modest, defensible billing improvement and a major workflow improvement. The workflow part is the bigger story; the billing part is a quiet upside.

---

For the patient safety side of the AI scribe documentation, see what to do when your AI scribe mishears a medication. For the malpractice and liability framing, what your malpractice carrier actually says covers the legal context.