E-News Exclusive

The Impact of Standardized AI Taxonomy on CDI

By Elizabeth S. Goar

In June 2026, the American Medical Association (AMA) CPT Editorial Panel revised Appendix S: AI Taxonomy for Medical Services and Procedures to improve AI reporting while preserving the framework’s core 2022 structure. Alongside temporary Category III CPT codes, the revisions created fundamental shifts in clinical documentation integrity (CDI) by reframing AI from an EHR tool to distinct, trackable clinical services with specific documentation and compliance requirements for billing and audit defense.

“It adds another layer or set of things that CDI needs to evaluate,” says Andrew Ibrahim, MD, MSc, chief clinical officer at Viz.ai. “If it’s potentially influencing care or tied to a billing code, then CDI needs to evaluate and ensure the documentation mirrors the care provided. In the case of AI, it would be both.”

Keeping Pace With AI
As AI capabilities expand and adoption accelerates, consistent terminology has become essential for aligning clinical oversight, payer policy, and regulatory expectations. Appendix S established a technology agnostic, output-based framework for classifying AI enabled medical services to guide how emerging AI innovations, ranging from expert systems and machine learning models to algorithmic services, are described within the CPT code set.

Real-world feedback revealed confusion around category boundaries and the need for more inclusive language capturing diverse AI architectures. This led to the most recent revisions, which:

• refine the distinctions between assistive, augmentative, and autonomous outputs;
• introduce formal definitions for key terms; and
• replace “machine” with “software output(s)” to reflect clinical reality.

“The original taxonomy gave us a foundation, but AI in clinical practice has moved quickly,” Richard Frank, MD, PhD, cochair of the Panel’s Digital Medicine Coding Committee and a member of the CPT Editorial Panel, shared in an AMA article explaining the changes. “The revisions sharpen the distinctions between categories in ways that reflect how the output from software is actually used by physicians to the benefit of patients today, and therefore also provide applicants with clearer guidance when they bring new applications to the Panel.”

Alongside the Appendix S revisions, the most recent Category III codes—temporary codes for emerging technologies to track utilization and establish clinical efficacy—took effect July 1, 2026. The update includes new codes for digital 3D modeling and analysis derived from patient imaging, as well as point-of-care quantitative analysis software tools. An additional 58 new codes covering therapeutic and surgical innovations will take effect on January 1, 2027.

Ibrahim notes that the updated taxonomy and temporary codes will impact CDI professionals in two ways. First, new AI tools are eligible for reimbursement with dedicated CPT codes that must be documented so CDI can verify that documentation and code usage align. Second, clinicians are increasingly using AI, adding another layer of verification for CDI professionals focused on documentation integrity.

“There is already a fair bit of variation among health systems in their CDI practices and what they feel comfortable with documenting and billing for,” Ibrahim says. “I suspect that variation will repeat itself with AI. Though, in general, I think there will be broader strong uptake of the AI codes as the ones approved all have quite clear value to patients.”

CDI Workflow Adjustments
Addressing the revisions requires CDI professionals to adjust their workflows to meet granular requirements, new documentation rules, and compliance oversight requirements. One of the biggest changes will be aligning CDI workflows with the taxonomy’s classification of software outputs by function and required clinician involvement. CDI workflows will also need to adjust in terms of:

• Focus: Rather than reviewing clinician narrative for missing diagnoses/specificity, under the AI taxonomy and Category III codes, CDI specialists must now validate clinician interaction with AI outputs and verify human oversight levels.
• Payer audit risk: Along with identifying and correcting unsupported severity of illness (DRG/CC/MCC), CDI specialists must now also look for lack of medical necessity or missing attestation for Category III AI codes.
• Query triggers: The need to clarify ambiguous clinician documentation now extends to confirming whether an AI output was accepted, contested, or overridden.

Documenting how a physician responds to AI output could increase query volume, but that shouldn’t negatively affect the CDI-clinician relationship, Ibrahim says. Tensions may arise if the tool’s output falls into the category of AI slop—when clinical notes appear to have been generated, heavily assisted, or copied by AI but add little clinical value and may contain generic, repetitive, inaccurate, or irrelevant content.

“I would say most clinicians find documentation a bit annoying and a distraction from the actual work of providing care to patients but accept that it is important to how the health system runs,” Ibrahim says. “Where friction could arise is if the chart turns into AI slop and there is a lot of extraneous information in the chart that is hard to verify or simply just way more volume. Then I could see some friction from CDI saying there is way more to verify here than a typical chart a year ago.”

Ultimately, while there have always been new codes to learn, AI tools bring with them a steep learning curve. For example, many of the AI codes encountered by CDI teams are likely to be related to EKGs and CTs. If the AI tool integrates with the EHR and feeds information back into it, CDI teams can more easily track the relevant data and verify how it was used.

“But many AI tools do not have EHR integration, so it can be harder for CDI to track down supporting data points to verify usage of the AI tool,” Ibrahim says. “If I were a CDI professional, I would want any AI tool that is tied to a CPT code to also have EHR integration so I could verify its use.”

Concerns Beyond Documentation
Beyond the mechanics of proving medical necessity, another real-world concern surrounding the rapid uptick in AI-augmented billing is the devaluation of the clinical chart. “From a clinical perspective, I am most worried about the chart just becoming a lot of noise,” Ibrahim says. “Already, since EHRs added dot phrases and templates, it's easier to pull more and more information into a note. Much of it just repeats what else is in the chart for completeness, making it hard to find what is new or essential in the note.”

He points to changing the order of the note as an odd but real development that has emerged over the past few years. Historically, a typical SOAP note would be ordered: Subjective, Objective, Assessment, Plan, he explains. But the S and O have gotten so long due to the many dot phrases and templates that “clinicians now put the Assessment/Plan right at the top. Kind of a ‘Give me the essentials up front.’ It makes you pause and think, ‘Is anyone reading the rest of the note? Is the rest of it just there for CDI?’”

Another concern is that AI tools might be used to expand the notes, pulling in more data and making them longer without adding value. “But I am optimistic about the opposite,” Ibrahim says. “What if the AI tools actually make the note far more concise and readable? That would be great for all groups involved.”

— Elizabeth S. Goar is a freelance health care writer based in Wisconsin.