AI Billing After CPT Appendix S

A discussion and analysis on:

Appendix S—Artificial Intelligence Taxonomy for Medical Services and Procedures

CPT 2025, AMA, Updated Dec 30, 2024

Appendix S—Artificial Intelligence Taxonomy for Medical Services and Procedures

Introduction

Appendix S in the 2025 CPT code set introduces a formal taxonomy for describing how artificial intelligence contributes to medical services and procedures. Rather than treating “AI” as a single concept, it focuses on the specific work done by machines on behalf of physicians and other qualified health care professionals (QHPs). The appendix classifies AI-enabled services as assistive, augmentative, or autonomous, with autonomous services further divided into levels based on how independently the system acts. This creates a shared language for discussing what the machine actually does in the clinical workflow and what level of human oversight is expected.

Detailed Analysis

1. Purpose and Scope of Appendix S

Appendix S is designed to guide classification of AI-driven services—expert systems, machine-learning models, and other algorithm-based tools—into assistive, augmentative, or autonomous categories. It emphasizes that “AI” is neither necessary nor sufficient to describe a service; what matters is how the procedure is performed and what portion of the work is delegated to the machine versus the clinician. The taxonomy is explicitly tied to clinical procedures and professional work rather than marketing language.

Why This Matters:  This shifts reimbursement and regulatory discussions away from vague “AI-enabled” claims toward concrete descriptions of machine work (detection, analysis, independent conclusions), which is what payers and regulators ultimately care about.

CareSight lens:  When evaluating an AI product, we should explicitly map the tool to an Appendix S category at the outset. That classification becomes the foundation for coding strategy, supervision requirements, documentation language, and risk framing.

2. Assistive and Augmentative AI Categories

Assistive AI refers to systems that detect or highlight clinically relevant data but do not independently analyze or interpret it. The clinician is responsible for reviewing the machine’s output, making clinical judgments, and documenting findings.

Why This Matters:  Both categories presume a human interpretive step, but augmentative tools shift more analytic workload to the machine, which can affect perceived value, code valuation, and expectations for documentation and quality oversight.

CareSight lens: For tools that flag findings, calculate risk, or summarize multimodal data, we should decide whether they simply assist detection (assistive) or truly produce new, clinically meaningful analysis (augmentative). That decision shapes whether an existing code pattern is appropriate and how we describe physician work relative to machine work.

3. Autonomous AI and Levels I–III

Autonomous AI describes services in which the machine independently interprets data and generates clinically meaningful conclusions without concurrent physician or QHP involvement. Appendix S further subdivides autonomous services into three levels:

• Autonomous Level I: The AI generates conclusions and recommended diagnosis/management options; implementation still requires physician or QHP action.

• Autonomous Level II: The AI draws conclusions and can initiate diagnostic or management steps, but a clinician has an opportunity to review and override.

• Autonomous Level III: The AI initiates management by default; clinician intervention is needed to contest or modify the action.

Why This Matters:  These levels formalize degrees of automation and clarify expectations for oversight, safety mechanisms, and liability. As tools move from Level I to Level III, scrutiny increases around governance, audit trails, and fail-safe behavior.

CareSight lens: Any product claiming to “automate” diagnosis or treatment should be placed on this spectrum. Where it lands will influence whether current codes are appropriate, whether new coding pathways may be needed, and how aggressive organizations can realistically be in their adoption timelines.

4. From Taxonomy to Codes: Early Signals in CPT 2026

The taxonomy is already surfacing in concrete codes. The CPT 2026 code set includes 288 new codes, with a significant subset devoted to digital health and AI-related services. These updates add: short-duration remote monitoring codes (2–15 days within a 30-day period, with new management codes at 10 minutes instead of 20) and multiple AI-augmented services such as coronary plaque assessment, perivascular fat analysis for cardiac risk, multispectral imaging for burn-wound classification, and algorithmic detection of cardiac dysfunction (AMA 9/2025).

Why This Matters:  These codes operationalize the Appendix S ideas: many of them describe algorithms that either help detect subtle findings or augment physician analysis, rather than fully replacing clinician judgment.

CareSight lens: When we see a proposed AI service, we can now ask whether it looks more like these early assistive/augmentative patterns—where codes already exist—or whether it is pushing toward higher-level autonomy that today’s code set only partially captures:

• If your tool’s behavior fits neatly into patterns for which CPT already has codes and language (e.g., “AI-augmented quantification that the physician interprets”), your billing story is usually easier.

• If your tool is more “autonomous” in a way that doesn’t map cleanly onto existing code patterns, there is potential to face a longer path to clear coverage, prior auth friction, etc.—which may impact reimbursement.

Practical Implications for Recommendations

Classifying AI Work Up Front:

In CareSight Feasibility Analyses, we should begin by assigning each tool a preliminary Appendix S category (assistive, augmentative, autonomous Level I–III) based on its real behavior in the workflow. This anchors subsequent discussions about coding options, supervision language, and documentation requirements.

Informing Coding and Reimbursement Strategy:

• Assistive and augmentative tools often align with existing patterns where a machine supports detection or analysis and a clinician remains responsible for interpretation and reporting.

• Autonomous tools—especially Levels II and III—may require stronger justification, more detailed safety narratives, and in some cases exploration of new codes or modifiers that explicitly account for machine-performed work.

Anticipating Governance and Adoption Questions:
The taxonomy provides health systems and payers with structured questions:

• What portion of the service is performed by the machine versus the clinician?

• When is clinician input required, and when is it optional or retrospective?

• How are overrides, disagreements, and failures tracked and resolved?

Our recommendations should prepare clients to answer these questions in policies, consent language, and implementation plans.

Concluding Reflection

Appendix S offers a pragmatic framework for describing what AI systems actually do in clinical services—whether they detect data, analyze and structure it, or independently generate and act on clinical conclusions. Integrating this taxonomy into feasibility work allows us to move beyond generic “AI-enabled” labels and instead ground assessments in machine role, oversight level, and likely coding and governance implications. Combined with the early wave of AI-related codes in CPT 2026, it signals that reimbursement conversations are rapidly shifting toward this more precise language. 

For teams building or deploying AI in care delivery, aligning their product descriptions with Appendix S is becoming less optional and more of a prerequisite for realistic billing, adoption, and risk planning.