MedicalCode AI is bringing the same retrieval-grounded architecture that powers our ICD-10 AI to CPT — outpatient and physician procedures, evaluation & management, surgery, radiology, pathology and medicine. CPT AI is in active build today; join the waitlist for early-access onboarding.
CPT (Current Procedural Terminology) is the code set used in the United States for outpatient and physician procedures, evaluation & management visits, anesthesia, and most non-inpatient services. It is maintained by the American Medical Association (AMA) and updated annually. Roughly 10,000 codes cover everything from a 15-minute office visit (99213) to a multi-vessel coronary artery bypass graft (33533).
CPT coding is harder for an AI than ICD-10 coding for three reasons: evaluation & management codes require a level-of-service decision (history, exam, medical decision-making — or time under the 2021+ guidelines) that depends on the entire encounter note, not a single span; CPT carries a dense network of modifiers (bilateral, multiple procedures, staged, distinct site, assistant surgeon) that interact with NCCI bundling rules; and the live CPT corpus is licensed by the AMA, so the AI ingests it under license rather than from a public source. None of these are architectural blockers — but they take CPT longer to ship at GA quality than ICD-10.
ICD-10 answers what is wrong with the patient. CPT answers what was done about it. Almost every U.S. claim carries codes from both systems. MedicalCode AI is shipping ICD-10 first (CM + PCS, live today) and CPT next.
| ICD-10 AI | CPT AI | |
|---|---|---|
| Coverage | ICD-10-CM (~74k codes) + ICD-10-PCS (~78k codes), live. | CPT (~10k codes) — in build. Join the waitlist for early access. |
| Used for | Diagnoses on every claim; inpatient procedures (PCS). | Outpatient and physician procedures, evaluation & management, anesthesia. |
| Coder time today | 15–30 min for moderate inpatient charts. | 5–15 min per outpatient encounter; level-of-service E/M is the bottleneck. |
| Maintained by | CDC/NCHS (CM) and CMS (PCS). | American Medical Association (AMA). |
Office visits, hospital E/M, consults, preventive medicine. The level-of-service decision (99202–99215 etc.) is the single hardest CPT call in primary care and the most commonly down-coded in audits. AI gets leverage here by reading the full encounter note rather than picking a level from a checkbox.
From skin biopsy to multi-vessel CABG. Surgery codes carry modifier-heavy logic — bilateral, multiple procedures, staged, distinct site, assistant surgeon — that map directly to spans in the operative report. AI can read the op note end-to-end and propose the procedure code plus the correct modifier set in one pass.
Diagnostic imaging, interventional radiology, radiation oncology. Contrast / no-contrast distinctions, professional vs technical components, and imaging-guidance bundling rules are mechanical once the report is read accurately — exactly where AI is strong.
Lab panels, surgical pathology, cytopathology, molecular pathology. High-volume, repetitive, rules-driven coding — a natural fit for AI throughput once the requisition and pathology report are parsed.
Cardiovascular procedures, neurology, ophthalmology, vaccines, infusions, dialysis, psychiatry. A mixed bag of specialty-specific rule sets — the AI handles each by retrieving the candidate code group and showing only those candidates to the model.
CPT AI is the use of large language models to read clinical documentation — office notes, operative reports, lab results — and produce the CPT (Current Procedural Terminology) codes a human coder would assign for outpatient and physician services. The "AI" part does the reading, code lookup, level-of-service reasoning, and modifier selection; a human coder still reviews and signs off.
Not yet. ICD-10-CM and ICD-10-PCS are generally available now. CPT is in active build — the architecture is the same (keyword extraction → group selection → code selection, grounded against the live CPT corpus), but CPT requires AMA licensing and additional level-of-service logic for E/M codes. Join the waitlist and we will email you when CPT AI opens for early access.
Three reasons. First, evaluation & management coding requires a level-of-service decision that depends on history, exam, and medical decision-making documented across the entire note — not a single span. Second, modifiers (bilateral, multiple procedures, staged, distinct site, assistant surgeon) interact with bundling rules. Third, CPT is licensed by the AMA, so the live code corpus must be ingested under license rather than pulled from a public source. None of these are blockers, but they take longer than ICD-10 to ship to GA quality.
No. CPT AI uses the same retrieval-grounded architecture as ICD-10 AI: the model is shown only candidate codes retrieved from an indexed copy of the live CPT corpus and is constrained to pick from that retrieved set. If a CPT code is in the output, it exists in the current CPT code book. Hallucinated codes are structurally impossible.
Yes — that is the design goal. The AI reads the full encounter note, scores history / exam / medical decision-making (or time, for the 2021+ E/M guidelines), and proposes the level-of-service code with the supporting documentation spans. As with everything else, the coder sees the reasoning chain and can override.
Yes. The same coding API that returns ICD-10 today will return CPT once the system is GA. We have integration paths for common outpatient EHR feeds (HL7 v2, FHIR, flat-file) and major RCM platforms.
Email info@medicalcodeai.com with "CPT AI waitlist" in the subject line and a one-line description of your use case (specialty, monthly encounter volume, EHR). We are onboarding early-access partners in cohorts and prioritising real outpatient and physician-services workflows.
We are onboarding CPT AI in cohorts. Email a one-line description of your specialty, monthly encounter volume, and EHR — we will reach out when your cohort opens.
