CMS proposes a MIPS activity for AI governance. Its example: AI-generated notes with clinician review.

If the clinicians you code behind have turned on an ambient scribe, CMS just wrote your read of that note into a MIPS activity. On July 16, 2026, CMS published its CY 2027 Physician Fee Schedule proposed rule in the Federal Register, and inside the Quality Payment Program section is a new improvement activity titled "Clinician Use of Artificial Intelligence (AI) to Improve Patient Care." One of CMS's own qualifying examples is using AI to "assist with documentation (e.g., AI-generated notes with clinician review)."

What CMS actually proposed

CMS proposes adding six new improvement activities beginning with the CY 2027 performance period and the 2029 MIPS payment year. Two go in the Care Coordination subcategory and four in Advancing Health and Wellness. The AI activity is the fifth of the six.

The description is short. It would let MIPS-eligible clinicians "implement and participate in organizational initiatives that improve patient care through the responsible and transparent use of artificial intelligence (AI) in clinical and operational workflows." The mechanism CMS names is governance, not usage: clinicians would "establish policies for evaluating and monitoring AI tools," or participate in developing and refining AI-enabled resources such as predictive analytics, clinical decision support, and risk-stratification models.

CMS points to Table F-B1 in Appendix 2 for the detail on each proposed activity.

The line to read twice: "with clinician review"

Here is the full example list, in CMS's words. The activity covers "using AI tools to summarize medical literature for clinical decision-making, assist with documentation (e.g., AI-generated notes with clinician review), support population health management by identifying care gaps, determine patient eligibility for clinical trials, or generate draft responses to patient questions for clinician review."

Two of those five end with a human checking the output. CMS didn't describe a tool that documents. It described a tool that drafts and a clinician who reviews. Put plainly, CMS is telling you the note isn't done when the tool stops typing, and that's the useful part of this proposal for anyone who reads notes for a living.

How solid is the 25% CMS cites

In a separate request-for-information section of the same rule, CMS writes that tools aimed at reducing clinician documentation burden, naming AI scribes, "have perhaps been the most widely taken up by clinicians," with "an estimated 25% penetration among all US physicians."

Treat that number as what it is. CMS attributes it to Doximity's State of AI in Medicine Report 2026, a vendor survey, not a CMS measurement, and it appears as context for an RFI rather than as an agency finding.

It's also worth being precise about what CMS did and didn't say. The scribe observation sits in an RFI several sections away from the improvement-activity proposal, and CMS never ties the two together. Read side by side, they point the same way: enough physicians are drafting with a scribe that the review step is worth writing down. That connection is our reading, not CMS's. It does track with what payers have already been doing to documentation that got denser after AI landed.

The six proposed activities, and where the coding desk touches them

The titles and subcategories are CMS's, quoted from the proposed rule. The right-hand column is our reading of where each one reaches a coding or HIM desk. It isn't CMS guidance.

The proposal credits the clinician. The coder finds the defect.

Improvement activities are a MIPS category, so the credit attaches to the MIPS-eligible clinician or group. That's worth naming plainly, because in most practices the person who notices that a drafted note flipped a negation, or carried a resolved condition forward as active, is the coder or the HIM reviewer reading it two days later.

CMS says clinicians would "establish policies for evaluating and monitoring AI tools." It doesn't say who does the monitoring, and the description doesn't publish a template. That gap is where a coding-QA program already lives. If your practice writes that policy, the monitoring loop it describes is the same note sampling a QA program already runs, and one artifact can serve both. A practice that already samples AI-drafted notes for the same defects a coder checks has most of the monitoring loop built.

What an AI-drafted note review has to catch

CMS asks for a policy to evaluate and monitor AI tools and doesn't say what to look for. Here's the list we'd start from. None of it is new to a coder. It's the same documentation reading you'd give any note, applied to a draft nobody has adopted yet.

This is our checklist, not a CMS one. The underlying discipline is unchanged: the record has to show the condition was evaluated and addressed on the date billed, and the evidence has to be in the note. A summary isn't the note. It's the same gap that drives denials for insufficient documentation. The note just gets written by something else first.

What isn't decided yet

This is a proposal. The activity IDs in the text are still placeholders, written as IA_AHW_XX, so there's no final identifier to build an attestation around. CMS is seeking comment on adding each of these activities beginning with the CY 2027 performance period and the 2029 MIPS payment year.

CMS is also proposing to modify five existing improvement activities on the same timeline, including IA_CC_8, "Implementation of documentation improvements for practice/process improvements." If documentation improvement is already how your practice attests, read that one too before it changes underneath you.

Comments must be received by September 14, 2026, and CMS asks commenters to reference file code CMS-1848-P.