Ambient AI Scribes Write Better Notes. They Can Also Write You Into an Audit.

Ambient AI scribes listen to the visit, then draft the note. For physicians buried in after-hours charting, that is the most useful thing AI has done in years — and adoption shows it. But a note is not just a clinical record. It is the evidence behind every code on the claim. Change how notes are produced and you change what gets billed, what gets audited, and what a payer’s algorithm sees when it decides whether your level 4 visit was really a level 4.

This is not an argument against ambient scribes. It is an argument for rolling them out with the revenue cycle and compliance team in the room, not just IT and the clinicians.

Three-quarters of physician AI users in the same survey said AI had already reduced their administrative workload. Multi-site research published in npj Digital Medicine in 2026 found ambient scribes cut EHR and documentation time — but by modest amounts that varied widely by specialty. The benefit is real; it is just not uniform, and it is not free.

A scribe captures everything said in the room. A human-written note captures what the physician had time to type. The scribe’s note is usually longer, more complete, and more likely to mention every problem discussed, every data point reviewed, and every risk factor — which is exactly the material E/M leveling and risk-adjustment coding are built from.

Some of that increase is legitimate: work that was always done but never documented. Some of it isn’t. A note that lists six problems because six topics came up in conversation doesn’t mean six problems were addressed in a way that counts toward medical decision making. Payers can’t tell the difference from claims data alone, so they respond to the pattern — and the response is what we’re already seeing in automated E/M downcoding programs.

Three specific risks to manage

• E/M level creep. If average E/M levels rise after go-live, you need to show the higher levels are supported by MDM or time — not by note length. Expect payers to notice the shift before you do.
• Cloned-looking documentation. CMS and its contractors have long scrutinized notes that look copied from visit to visit. Templates plus AI drafting can produce strikingly similar text across encounters; each note must reflect that specific visit.
• Risk-adjustment exposure. In Medicare Advantage and other risk-adjusted contracts, diagnoses captured in a note can raise risk scores. A diagnosis must be supported as monitored, evaluated, assessed, or treated at the visit — a mention isn’t enough, and unsupported HCC capture is a recurring federal enforcement theme.

• Recording consent. About a dozen states — including California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania, and Washington — require consent from everyone in a conversation before it is recorded. Get explicit consent, document it, and give patients a simple way to opt out.
• Business associate agreement. The scribe vendor receives audio and PHI, which makes it a business associate under HIPAA. Sign the BAA, and ask which sub-processors and model providers touch the audio — each needs to be covered. Our guide to generative AI and PHI covers the questions to ask.
• Retention. Know whether audio is kept, for how long, and whether it is used to train the vendor’s models. “Deleted after the note is generated” and “retained for product improvement” are very different contracts.
• Authentication. CMS requires medical records to be authenticated by the treating provider. The signature means the provider reviewed and agrees with the note — including anything the AI wrote.