Artificial intelligence has already found its way into the pathology and laboratory space, and the coding rules that govern it are catching up. Recent revisions to CPT® Appendix S, paired with a proposed Medicare payment shift, could change how your lab reports and gets paid for AI-driven testing — and it’s worth understanding both pieces before they take effect.
Appendix S isn’t a code list. You won’t find a lookup table of every AI-enabled 80000-level or proprietary laboratory analysis (PLA) code inside it. Instead, it’s a taxonomy — a framework for classifying how a qualified healthcare provider (QHP) and software work together on a given procedure, and how much of the diagnostic or management decision each one is responsible for. That framework has ripple effects well beyond terminology: it shapes future code descriptors, valuation, coverage decisions, regulatory scrutiny, and even how labs structure their workflows and business models. This isn’t the first time CPT® has had to build new structure around emerging technology; coders who followed how the code set absorbed augmentative AI into cardiac imaging or tracked the steady stream of new Category III codes for emerging services will recognize the pattern here.
What Changed on June 8, 2026
The AMA’s CPT® Editorial Panel formally accepted revisions to Appendix S, the product of roughly four years of stakeholder feedback aimed at sharpening the lines between the existing AI categories. Appendix S still doesn’t attempt to define “AI” outright — instead, it describes the kinds of applications and software that fall under the umbrella, things like expert systems, machine learning models, and algorithm-driven tools used for diagnosis or management. With that groundwork in place, the revised taxonomy sorts procedures into three tiers based on the software’s output and its role in patient care.
Assistive. The software surfaces clinically relevant data but doesn’t derive a parameter or provide an interpretation — that step still belongs to the QHP. Assistive output can support the provider’s own performance (accuracy, precision, reducing variability between observers) without needing to independently meet a “clinically meaningful” bar on its own. Typical language here includes phrases like “likelihood of,” “suggestive of,” or “risk for.”
Augmentative. Here the software produces a quantitative or categorical parameter that goes beyond simple calculation on the input data — think an index, a categorical classification, or a risk score. To qualify as augmentative, that output has to be clinically meaningful, distinct from the raw inputs, and validated against real clinical outcomes. Language like “predictive of” or “prognostic of” often signals this category. Augmentative procedures still stop short of providing an interpretation; that’s left to the QHP, sometimes folded into a separate service like an E/M visit that uses the result.
Autonomous. This is where the software takes over more of the clinical judgment, and CPT® further splits it into three levels based on how much the QHP remains in the loop:
- Level I — The software recommends a specific diagnosis or intervention based on derived parameters, but the provider still decides whether to accept or reject it.
- Level II — Similar to Level I, except the algorithm can initiate an action on its own. The QHP gets an alert and has the opportunity to stop it before it goes through.
- Level III — The most automated tier: the software starts medical management based on its own conclusions, and while the QHP provides oversight, the action proceeds unless the provider actively intervenes.
Where This Shows Up in Real Codes Today
Most AI-enabled pathology and lab codes on the books right now fall into the assistive or augmentative categories, not autonomous. A few familiar examples:
- 0003M — Liver disease panel using ten biochemical assays (ALT, A2-macroglobulin, apolipoprotein A-1, total bilirubin, GGT, haptoglobin, AST, glucose, total cholesterol, and triglycerides), reported as a prognostic algorithm generating quantitative scores for fibrosis, steatosis, and NASH.
- 0376U — Prostate cancer image analysis of at least 128 histologic features plus clinical factors, producing a prognostic algorithm for distant metastasis risk and prostate cancer-specific mortality, with a predictive component for androgen deprivation therapy response where applicable.
- 81546 — Thyroid oncology mRNA gene expression analysis across 10,196 genes from a fine needle aspirate, reported as a categorical result such as benign or suspicious.
Coders who work regularly with molecular pathology and genomic sequencing codes will recognize the general structure of these descriptors — it follows the same pattern seen when new targeted genomic sequence analysis codes were added in recent cycles.
Key Definitions Worth Memorizing
A few terms anchor the entire Appendix S framework:
- Parameter (derived): A quantitative or categorical output — an index score or classification — that results from more than a straightforward mathematical calculation on the input data.
- Clinically meaningful: Documentation showing the output genuinely contributes to patient management — diagnosis, treatment, mitigation, cure, or prevention of a condition.
- Automated: The algorithmic process of deriving parameters from inputs, where those inputs may themselves come from a separately performed service.
AI’s Reach Extends Beyond Diagnostic Codes
The codes above only capture AI used directly in patient-facing diagnostic work, such as image analysis for histopathology and microbiology or gene expression analysis for cancer. Labs are also deploying AI behind the scenes — flagging and discarding damaged specimens, automating routine lab processes, running quality control checks, and forecasting testing demand to manage staffing and supply. These applications can shorten turnaround time and cut down on human error, but broader adoption is still running into real obstacles: cost, validation requirements, interoperability between systems, regulatory uncertainty, and ethical concerns around privacy and algorithmic bias.
The Payment Question Coders Should Watch Closely
Here’s where this shifts from a coding classification exercise to something with direct reimbursement consequences. Many AI-driven lab and pathology services are priced based on software licensing costs rather than the instrumentation and staff time that traditionally set lab test pricing. That distinction matters more every year as AI use grows, because it changes the cost basis payers are actually reimbursing against.
CMS picked up on exactly this distinction in its proposed rule for the 2027 Medicare Physician Fee Schedule (MPFS). The proposal draws a line between traditional lab tests that generate brand-new results and software-based analytics that apply proprietary algorithms to data that’s already been generated — CMS is calling this second category “software as a medical service” (SaMS). Under the proposal, SaMS codes would move away from the Clinical Laboratory Fee Schedule (CLFS) and be paid instead under the MPFS or the Outpatient Prospective Payment System (OPPS). Nine lab tests currently priced under the CLFS are named in the proposal for this shift, including 0376U from the list above.
For coders, this isn’t just a billing department concern. A code moving from CLFS to MPFS/OPPS pricing can change everything from expected reimbursement to how the service interacts with other same-day billing — a useful refresher here is how payment status indicators determine whether a code is paid separately or bundled, since that logic underpins exactly this kind of fee schedule migration. If your lab bills any of the affected codes, it’s worth flagging this proposal to your billing team now rather than after the final rule lands.
Bottom Line
Appendix S doesn’t hand coders a new set of codes to memorize, but it does something arguably more important: it gives everyone — coders, payers, and the AMA itself — a shared vocabulary for how much of a diagnostic decision came from software versus a human provider. As more PLA and MAAA codes get built or revised around this taxonomy, and as CMS experiments with separating AI-based analytics from traditional lab pricing, understanding where a given test falls on the assistive-to-autonomous spectrum will matter for code selection, documentation review, and reimbursement forecasting alike.
