Key points
- CMS states that the volume of documentation should not be the primary influence on the level billed, so the code follows the decision making and time recorded.2
- When OIG coders reviewed 2010 Medicare E/M claims, 42% were coded incorrectly in one direction or the other, so accuracy means catching undercoding as well as upcoding.3
- Score any scribe on seven tests, run on 50 of your own visits against an independent certified coder, before you sign.
- wai suggests codes from the note, the clinician reviews and signs, and the claim and the audit log sit on the same record as the note.678
Summary
Judge an AI scribe on whether its suggested codes match what the signed note supports, in both directions. Test it on 50 of your own visits against an independent certified coder, track code mix and denials by clinician for 90 days, and make sure the clinician reviews every code before it is billed.
Each visit is billed at the level the clinician delivered, and every code can be traced to the signed note behind it. That is the standard to hold an AI scribe to, measured on your own visits before you sign. The gain worth having comes from a more complete note, never from a longer one.
See it in your clinicThirty minutes on your own patients and doctors.
Request a demoSee a consultation become a signed note with suggested codes and the reason for each.
| Test | What good looks like | Red flag | Weight |
|---|---|---|---|
| Code match on 50 of your visits | Suggested codes match an independent certified coder on most visits | An accuracy figure measured on other clinics' visits | High |
| Direction of mismatches | Misses split both ways with no steady drift | Most misses sit one level above the coder | High |
| Reason for each code | Each code points to the decision making or time in the note | A code with no link back to the note | High |
| Clinician sign-off | Codes stay suggestions until the clinician approves them | Codes flow to the claim without review | High |
| Time-based visits | Total time entered or confirmed by the clinician | Time inferred from the length of the recording | Medium |
| Audit log | Drafts, edits, code changes and signature logged with times and exportable | No record of what the clinician changed | Medium |
| Denials after go-live | Denial rate flat or falling across 90 days | Rising denials for medical necessity or documentation | Medium |
Coding is where a scribe earns its return
The return a scribe shows in the accounts comes from the codes billed on each visit. A note that records the decision making and the time spent supports the level of service the clinician delivered, and a fuller note supports a higher code wherever the work was done.
What the evidence shows so far
The clearest data comes from UCSF Health. Researchers compared 698 physicians who adopted an ambient AI scribe with 867 who did not, across ambulatory visits from January 2023 to April 2025. Adopters billed 1.81 more RVUs a week and saw 0.80 more patients a week, and there was no difference in the proportion of claims with a denial.1
The authors put the gain at about $3,044 a year per physician at the 2025 Medicare Physician Fee Schedule, and call for research on whether higher RVUs reflect more clinical services or more accurate coding.1 It is one site with an observational design, so measure your own result.
The rule every suggested code must pass
For Medicare in the US, the standard is written into the claims manual. Since January 2023, most office visits are levelled on medical decision making or on the clinician's time. CMS states that medical necessity is the overarching criterion for payment, that billing a higher level when a lower one is warranted is not appropriate, and that the volume of documentation should not be the primary influence on the level billed.2
Long, thorough notes read like complex visits, so the code follows the decision making and the time the clinician recorded.
When OIG had certified coders review a random sample of 2010 Medicare E/M claims, 42% were incorrectly coded, including both upcoding and downcoding, and 19% lacked documentation.3 Medicare's own error testing still put the fee-for-service improper payment rate at 6.55% for fiscal year 2025, and 8.44% for Part B providers.5 Undercoding is revenue that no audit returns, as set out in our guide to undercoding.
The audit trail decides the outcome
Auditors already look at how electronic notes are produced. OIG found that not all Medicare contractors could tell whether a provider had copied language or overdocumented in an electronic record, and recommended that CMS direct contractors to use providers' audit logs.4
The log is the clinic's evidence. It should record what was drafted, what the clinician changed, which code was suggested, which was billed and who signed, with times. When a payer asks why a visit was billed at a given level, the clinic shows the clinician's note and the code that followed it.
How to test a scribe on your own visits
Take 50 recent visits across your main visit types, run each through the scribe, and have a certified coder who has not seen the suggestions code each signed note independently.
Compare the match rate and the direction of every mismatch. Upward drift is a compliance risk, and downward drift leaves earned revenue unbilled. The scorecard below sets the bar.
The note, the code and the claim on one record
wai Clinicians drafts the note during the consultation and suggests codes from it, and the clinician reviews, edits and signs.6 wai Admin prepares the claim and the invoice from the same signed note for the team to approve, and every draft, approval and signature is written to an event log the clinic can export.78 The code an auditor sees is the code the clinician signed, with the note beside it.
Where to start
Pull last quarter's code mix by clinician and visit type, and the denial rate by reason. That is the baseline. Run the 50-visit coder test on every scribe on the shortlist and score each one on the table above.
In the pilot, keep a certified coder reviewing a sample every week, track code mix and denials by clinician for 90 days, and look into any clinician whose level distribution moves sharply in either direction. Our guide to reducing claim denials covers the denial side of the baseline.
Bring ten de-identified visits. We show the note, the suggested codes and the reason behind each, so your coder can check them line by line.
Frequently asked questions
Do AI scribes cause upcoding?
How many visits should we test?
Fifty recent visits across your main visit types is enough to see the match rate and the direction of misses. Have a certified coder code the signed notes without seeing the scribe's suggestions.
Who is responsible for the code that is billed?
The billing clinician and the clinic. The scribe suggests and the clinician signs, which is why every suggested code should be reviewed before the claim is submitted.
Will an AI scribe raise our revenue?
It can where notes currently miss work that was done. At UCSF Health the gain was 1.81 RVUs a week per adopting physician, about $3,044 a year at 2025 Medicare rates.1 Measure your own baseline first so the change can be attributed.
What should an auditor be able to see?
Who drafted, edited and signed the note, when, and which code was billed from it. OIG has recommended that Medicare contractors use providers' audit logs when reviewing electronic records.4
Sources
- 1JAMA Network Open, Holmgren and colleagues, 2026Ambient Artificial Intelligence Scribes and Physician Financial Productivity (open-access full text in Europe PMC)Read 14 October 2026Read source
- 2Centers for Medicare and Medicaid Services, 2026Medicare Claims Processing Manual, Chapter 12, section 30.6.1, selection of level of evaluation and management serviceRead 14 October 2026Read source
- 3HHS Office of Inspector General, 2014Improper Payments for Evaluation and Management Services Cost Medicare Billions in 2010 (OEI-04-10-00181)Read 14 October 2026Read source
- 4HHS Office of Inspector General, 2014CMS and Its Contractors Have Adopted Few Program Integrity Practices To Address Vulnerabilities in EHRs (OEI-01-11-00571)Read 14 October 2026Read source
- 5Centers for Medicare and Medicaid Services, 2025Comprehensive Error Rate Testing (CERT), fiscal year 2025 improper payment rateRead 14 October 2026Read source
- 6wai, 2026wai for clinicians, notes, letters, coding and cited evidenceRead 14 October 2026Read source
- 7wai, 2026wai Admin, the front desk and back officeRead 14 October 2026Read source
- 8wai, 2026Security at waiRead 14 October 2026Read source
Bill every visit at the level you delivered, with the note to show for it.
Fifteen minutes on your own visit types, with every code tied to the signed note. We reply within one working day.
Partner with us
See it on your own patients, then build it with us
Thirty minutes on your own patients, doctors and rules, in your market. We reply within one working day, and the clinics that partner with us early shape what we build next.
See what wai does between the visits