Summarize this blog post with:
ChatGPT
Perplexity
Claude
GrokThe short answer: AI scribes can suggest an E/M level and diagnoses, but they do not bill, and the evidence on upcoding is unsettled. Trilliant Health found the share of level 4 and 5 established-patient visits rose by 7 to 12 percentage points between 2018 and 2024 across six systems that announced AI scribing, but the study could not isolate the scribe's effect. Audit your own notes.
If you run a practice, the question behind the headline is simple: will this tool change my billing, and will I be able to defend it? Here is what the data shows, what it cannot show, and a method to check your own.
What claims data shows
Trilliant Health analyzed national all-payer claims for six large systems that publicly announced ambient AI scribing. The table shows the share of visits billed at level 4 or 5 (99214 and 99215 for established patients, 99204 and 99205 for new patients).
| System | Established, 2018 | Established, 2024 | Established, change | New patient, 2018 to 2024 | New patient, change |
|---|---|---|---|---|---|
| A | 59.8% | 67.2% | +7.4 pp | 51.3% to 64.8% | +13.5 pp |
| B | 40.9% | 52.7% | +11.8 pp | 42.9% to 57.6% | +14.7 pp |
| C | 65.3% | 72.9% | +7.6 pp | 60.5% to 80.0% | +19.5 pp |
| D | 50.4% | 60.2% | +9.8 pp | 44.5% to 63.7% | +19.2 pp |
| E | 56.6% | 64.6% | +8.0 pp | 47.5% to 67.0% | +19.5 pp |
| F | 47.8% | 57.7% | +9.9 pp | 43.7% to 60.2% | +16.5 pp |
Source: Trilliant Health, claims data 2018 to 2024.
What this does not prove
- No comparison group. There was no set of non-adopting systems, so the rise cannot be pinned on scribes.
- Selection. The six systems were chosen because they announced adoption.
- Rule changes. The authors tie the timing partly to the 2021 E/M revisions, which let office visit levels be chosen by medical decision making or total time instead of history and exam (see the AMA's CPT E/M revisions FAQs).
- Interpretation. The authors say the causes are not clearly understood and that the rise likely reflects more complete, rules-based documentation. Whether that is accurate depends on comparing notes with true clinical complexity, which the study did not do.
The policy view
In a January 2026 JAMA Health Forum viewpoint, Nong and Neprash put scribes at roughly $300 to $500 per clinician per month and noted, as an illustration, that a $500 monthly cost could be recouped by about four extra level-4 established-patient Medicare visits per month. They warn that using scribes for billing, not just documentation, could raise spending, and they recommend clinician-in-the-loop claims auditing. Expect payers to look harder at this. For what the time and burnout studies found, see do AI scribes save time.
What this means for your practice
A scribe changes the note. The note drives the level. So your compliance risk is the note, whoever or whatever drafted it. Billing a level the signed note does not support is the problem in every case.
The athenahealth AI scribe, athenaAmbient, describes its notes as "ready for coding" and does not describe automated code assignment. SteerNotes shows E/M level and diagnosis suggestions with the supporting evidence inline, and requires the clinician's signature. Evidence beside each suggestion makes the audit below faster.
A 30-day audit
- Pull 20 random signed notes per clinician in the first month. You learn the baseline across visit types.
- Record the tool's suggestion, the clinician's final level and a coder's independent level. You learn where they diverge.
- Log the direction of each miss: too high or too low. You learn compliance risk versus lost revenue.
- Calculate the override rate per clinician. You learn whether suggestions are trusted.
- Review every note where the clinician accepted a level the coder rejects. You learn your real exposure.
- Repeat at day 60 with a fresh sample. You learn whether accuracy holds.
Step 5 is the number to watch. An override is the system working. An accepted suggestion the coder disagrees with is what costs money later. The sample of 20 is a starting point, not a standard.
Questions to ask any vendor
- Which code sets and visit types are suggested? Many tools cover office E/M only.
- Is the evidence for each suggestion visible in the note? You cannot audit what you cannot see.
- Can it support both decision-making and time-based levels? Both methods are valid.
- Are clinician changes logged? You need an audit trail.
- Is anything submitted without a signature? That is autonomous coding risk.
Next step
See coding suggestions with the evidence shown beside them. Book a SteerNotes demo.
