Match the code to the category
Then read the guideline
- Read the CPT guideline for the code family before you build the charge
- The descriptor is written in Appendix S language on purpose
Guide · Specialty coding
CPT 2027 AI codes: 43 codes, three categories, one question to ask of each
Short answer: CPT 2027 adds 10 AI-related codes for a total of 43, all organized by the Appendix S taxonomy. Ask one question of every AI tool: does a clinician have to read the output before it means anything? If yes, it is assistive and the visit is what you bill. If no, find the code that describes the software's own work.
Released Sep 9, 2026 · effective Jan 1, 2027
Appendix S · assistive, augmentative, autonomous
Category III · T codes, no national value
10 of 453 · CPT 2027 changes
Focus
CPT 2027 AI codes
Metric
43 codes
Framework
Appendix S
Codes
Category III
Effective date
Jan 1, 2027
Length
6 min read
10new
AI-related codes added in CPT 2027, released September 9, 2026 and effective January 1 [1]
43total
AI-related CPT codes after the 2027 additions [1]
3categories
Appendix S: assistive, augmentative and autonomous (Levels I to III) [2]
453changes
Total CPT 2027 changes; the AI codes are 10 of them [1]
The release gives the structure, not the code numbers
The AMA's September 9, 2026 release adds 10 AI-related codes and updates the Taxonomy for Artificial Intelligence in Medical Services and Procedures, the document that tells code applicants which words to use in a descriptor [1]. It does not list the 10 by number or section, so this post does not either. Every AI service now sits in one of three Appendix S categories, and the category decides how the service is billed [2].
Category 1Assistive
Output provides clinically relevant data without deriving a parameter or generating an interpretation; may be a likelihood or risk estimate.
Clinician roleRequired: physician or QHP interpretation and report
Category 2Augmentative
Output is a quantitative or categorical parameter qualitatively different from the input, beyond summation, averaging or descriptive statistics.
Clinician roleOptional: may adjust settings; work captured by existing codes such as E/M
Category 3 · Level IAutonomous I
Software derives parameters and independently generates clinically meaningful interpretations; recommends a diagnosis or specific management.
Clinician roleDecides whether to implement or reject
Category 3 · Level IIAutonomous II
As Level I, and recommends medical management with a reasonable opportunity to stop the action, for example through an alert.
Clinician roleCan halt the action before it is carried out
Category 3 · Level IIIAutonomous III
Software automatically starts management actions that continue unless a physician intervenes.
Clinician roleOversight and performance review across multiple interventions
Category to billing consequence
| Category | Clinician role | Billing consequence |
|---|---|---|
| Assistive | Required: the output requires physician or QHP interpretation and report | The physician service is the billable event; the software supports it |
| Augmentative | Optional: the physician may adjust settings | Look for a code that describes the software analysis itself; the visit is billed as it always was |
| Autonomous, Level I | Decides whether to implement or reject the recommendation | The software performs the interpretive work the code describes |
| Autonomous, Level II | Can halt the action before it is carried out | Same, with the stop point documented |
| Autonomous, Level III | Oversight and performance review, typically across multiple interventions | Same, with oversight documented across the series |
AMA, CPT Appendix S [2], definitions paraphrased
Takeaway
Ask one question of every AI tool in the building: does a clinician have to read the output before it means anything? If yes, it is assistive and the visit is what you bill. If no, find the code that describes the software's own work.
Three operating rules from the descriptor
A vendor demo rarely says where its product sits in Appendix S; the code it suggests you bill does. If the descriptor says software analyzes and the physician interprets, the money on the claim is still the E/M or the imaging interpretation. If it describes software generating the interpretation, the physician's separate interpretation of the same finding may not be separately reportable.
Do not stack an interpretation on an autonomous service
Unless the guideline allows it
- The point of the autonomous category is that the software did the interpreting
Keep the software's output in the record
Documentation
- An assistive or augmentative service without the data it produced is a billed service without documentation
Payer check first, charge build second
Many AI services entered CPT as Category III tracking codes ending in T, with no national Medicare relative value; payment is set by each payer or priced by the contractor. A Category I code has a national value once CMS finalizes it in the Physician Fee Schedule. The AMA release does not say how many of the 43 are Category I versus III, so treat every code on a vendor's list as unpriced until the payer's policy is in hand.
| Check | Where it comes from | What to record |
|---|---|---|
| Appendix S category | The code descriptor and guideline in CPT 2027 | Assistive, augmentative or autonomous level |
| Category I or III | The code format and the CPT 2027 code list | Whether a national value exists |
| Payer coverage | Each payer's medical policy for the code | Covered, non-covered, or prior authorization required |
| Documentation | The software output and the clinician's note | What has to be in the record for the category |
| Scrubber rule | Your claim editor | Which combinations with E/M or interpretation codes are blocked |
The AI codes are 10 of 453 changes; run the crosswalk once
Named blocks in the CPT 2027 release, code count
Imaging and cardiology departments adopting AI tools are often the ones touched by the ventricular assist device and radiology changes [1]. Section table: The Coding Operator on the 453 changes; largest deletion block: the maternity transition playbook.
Maternity codes deleted
173.8% of 453
Maternity codes added
122.6%
AI-related codes added
102.2%
Diaphragmatic hernia repair codes added
92.0%
Unattended sleep study codes added
61.3%
Left ventricular assist device codes added
30.7%
DeletedAddedAI codesAlso: updated prostate biopsy codes and a time-based biofeedback structure
Two things share the phrase; only one is judged by accuracy
AI codes are CPT codes for AI-performed clinical services; coding AI is software that assigns codes to a chart. Appendix S describes the first. The second is judged by accuracy, reported honestly by category, not as one number. Our autonomous coding capability publishes a day-one auto-accept rate and a tuned rate because the two differ; the coding engine page describes how each rule pack is refreshed against a CPT release.
ASC surgical coding pilot, accuracy by category on the client's QA sample
Pilot-stage, one client. The weak category is the one we publish, because that is the one a coder has to watch.
CPT selection
95.3%strongest category
Modifiers
88.9%
Overall, upgraded engine
86.1%up from 79.5%, +6.6 points
Overall, old engine
79.5%baseline
Secondary diagnoses
53.5%the category a coder watches
0%25%50%75%100%
Above 90%80 to 90%Overall, upgradedOverall, oldBelow 60%
Four checks before any AI code reaches the charge master
List every AI tool in the buildingClassify each against Appendix S from the code descriptor, not the vendor demo.
Pull the payer policy for each codeCategory III codes are unpriced until a policy says otherwise.
Write the scrubber ruleBlock interpretation codes stacked on autonomous services unless the guideline allows them.
Ask vendors the same three questionsAppendix S category, payer policy, what the output looks like in the record. The multispecialty group billing team runs that review in any new-service build.
Frequently asked questions
How many AI codes are in CPT 2027?
The AMA says CPT 2027 adds 10 AI-related codes for a total of 43. The release does not list them by number; they are organized by the Appendix S Taxonomy for Artificial Intelligence in Medical Services and Procedures, which the AMA updated with this release.
What is the difference between assistive, augmentative and autonomous AI in CPT?
Assistive software provides clinically relevant data that still requires physician or QHP interpretation and report. Augmentative software produces a quantitative or categorical parameter that is qualitatively different from the input, with physician interaction optional. Autonomous software derives parameters and independently generates clinically meaningful interpretations, at Levels I to III depending on how much the physician can intervene.
Can I bill a physician interpretation on top of an autonomous AI code?
Check the CPT guideline for that code family first. The autonomous category describes software that performs the interpretive work, so a separate physician interpretation of the same finding may not be separately reportable. Document what the clinician did beyond the software output.
Are AI CPT codes paid by Medicare?
Only Category I codes carry a national value once CMS finalizes it in the Physician Fee Schedule. Category III codes, which many AI services use, are tracking codes without a national value, so payment depends on each payer's policy or contractor pricing. Confirm coverage before building the charge.
- AMA press release, AMA releases CPT 2027 code set (September 9, 2026): 10 new AI codes, 43 total, taxonomy update, section highlights
- AMA, CPT Appendix S: Taxonomy for Artificial Intelligence in Medical Services and Procedures (assistive, augmentative, autonomous)
Checked October 25, 2026. Rules change; confirm against the source before relying on them. Items marked proposed are not final until the final rule publishes.
Adding an AI service in 2027?
We classify it against Appendix S, pull the payer policies, build the charge and scrubber rules, and show you what the record has to contain before the first claim goes out. Free 30-minute coding review with a senior partner.