CPT 2027 sharpens the AI taxonomy. How a tool's role is classified decides how it's coded.

At its May 2026 meeting, the American Medical Association's CPT Editorial Panel accepted revisions to Appendix S, the framework that sorts AI-enabled services into three buckets, for the CPT 2027 code set. The edit is small in word count and large on your desk, because Appendix S is what sorts an AI-assisted service into a reporting category, or rules that it isn't a reportable service at all, and the deciding factor is the one you read the note for.

What Appendix S actually sorts

Appendix S doesn't rate a tool by accuracy, brand, or how impressive the demo looked. It classifies the service by the work the machine performs on behalf of the physician or other qualified health care professional. That gives three categories, assistive, augmentative, and autonomous, and the line between them is how much of the clinical judgment the software did and how much the clinician still owns.

The AMA says the May 2026 revisions "clarify and strengthen the taxonomy" and add "a more precise definition of what constitutes a clinically meaningful output," drawing on four years of real-world feedback from applications submitted under the existing appendix. Read plainly, the Panel tightened the seams between the three buckets so the same service lands in the same category no matter who submits it.

The three buckets, and the oversight each one demands

Here's the taxonomy in the terms that reach a claim. The middle column is a short quote from the appendix. The right two columns are our reading of what the record has to show and what you confirm before the service is billed. Those aren't AMA instructions.

The pattern is the same one you already apply. A code follows the work, and the work has to be in the note. A tool that only surfaces data for the clinician to read is a different billable event than a tool that reaches its own conclusion.

Autonomous got the sharpest edit

The autonomous category is where the classification does the most work, because it splits into three levels by exactly how much a human is in the loop. The AMA's own language, quoted:

None of these levels removes the clinician. They move where the clinician sits relative to the software's action, and that position is what the code is priced against.

A code existing is not the same as getting paid

The first CPT code for autonomous AI is already live and worth studying, because it shows the whole idea in one line. Code 92229 covers "imaging of retina for detection or monitoring of disease; point-of-care autonomous analysis and report." The software captures the image, analyzes it, and issues a report without the clinician interpreting it, which is what puts it in the autonomous bucket rather than the assistive one that governs a clinician-read retinal photo.

Two cautions travel with that. First, most AI in a clinic never earns a code. Appendix S covers AI work that maps to a reportable clinical service, so ambient scribes, schedulers, and inbox drafters are workflow software, not a reportable service, and generally carry no CPT code, which is why the MIPS improvement activity CMS proposed for AI governance is about policy and monitoring, not a fee. Second, a code existing doesn't make the service covered. A CPT code is a way to describe work, and payer coverage still varies, so the presence of a number isn't a green light to bill without checking the plan.

Where the human coder still sits

The uncomfortable version of "autonomous" is the one leadership repeats: the machine does it, so the coder is optional. The taxonomy says the opposite. Every category, including autonomous, is defined by a physician-oversight requirement, and someone has to confirm that requirement is in the record before the claim goes out. The vendor's label is only a starting point; the note decides. If a tool is marketed as autonomous but the note shows the clinician interpreted the output, the service is assistive, and the code changes with it.

That confirmation is a coding read, not a software feature. You're checking that the work described matches the category billed, that the oversight the level requires is documented, and that the evidence is actually in the note rather than assumed from the tool's marketing. It's the same discipline that keeps AI-drafted documentation honest before it becomes a code, pointed at the billing question instead of the diagnosis question.

What to do with this before January

This is a code-set revision, not a rule with a comment window, so there's nothing to file. There's something to learn. Before the CPT 2027 set is in front of you, get familiar with the three categories and the autonomous levels, because the classification will start showing up in payer edits and audit questions the moment AI-assisted services carry these codes. When a note names an AI tool, the category it belongs to, and the oversight behind it, is a question you can already answer from the official code language and the record in front of you.