HCC risk adjustment AI scribe 2026: Medicare Advantage coding that actually captures the dollars

May 8, 2026 · 8 min read

Hierarchical Condition Categories (HCCs) drive Medicare Advantage payment. Each captured HCC adds $1,500–$5,200 to a member's annual capitation, and the 2026 CMS-HCC v28 model is now fully phased in — meaning the diagnoses that mattered five years ago no longer count, and the diagnoses that count now require sharper documentation than most primary-care AI scribes are tuned for.

If you're a clinician seeing 20+ MA patients per day on a population health contract, undercoded HCCs are the single largest line of revenue you can recover without seeing more patients. Here's what actually moves the number in 2026.

The v28 model in one paragraph

CMS-HCC v28 (phased in 33% / 67% / 100% across 2024–2026) reduced the model from 86 HCCs to 115 categories with stricter clinical specificity. Diabetes without complications dropped from RAF 0.105 to 0.166 (renormalized but harder to capture passively). Vascular disease was restructured. Mental health and substance use HCCs moved upward. Net effect: the same patient panel codes ~3.5% lower under v28 unless documentation gets more specific.

Where AI scribes built for SOAP notes fail at HCC

A general-purpose AI scribe captures the encounter narrative. An HCC-aware scribe captures the encounter plus the supporting documentation that survives a RADV audit.

The $1,500–$5,200 math, per patient per year

Risk adjustment payment = base capitation × RAF (Risk Adjustment Factor). The 2026 PMPM benchmark in most counties runs $1,000–$1,400. Each 0.1 RAF point ≈ $1,200–$1,700 in annual revenue per member.

Common HCCs and their v28 RAF values:

• HCC 36 — Diabetes with chronic complications: 0.302 (~$3,600/yr)
• HCC 226 — Heart failure unspec: 0.331 (~$3,900/yr)
• HCC 263 — Stroke late effects with hemiparesis: 0.224 (~$2,600/yr)
• HCC 138 — Major depression severe single episode: 0.309 (~$3,700/yr)
• HCC 21 — Morbid obesity (BMI ≥ 40 + comorbidity): 0.250 (~$3,000/yr)
• HCC 32 — Drug/alcohol psychosis: 0.484 (~$5,800/yr)

A panel of 250 MA patients with 1.5 missed HCCs per patient per year — very common in unaudited primary-care — is $750k–$1.2M of unrecovered revenue annually for a small group.

What an HCC-aware scribe pipeline actually does

1. Pre-visit prefetch. Pull problem list, last 12 months of dx codes, recent labs/imaging, current med list. Compute "suspect HCC list" — conditions billed in past 12 months but not yet in current year, plus conditions implied by labs (eGFR, A1c, BMI) but not documented.
2. In-visit transcription. Whisper-class ASR captures the full encounter audio. Cost matters here: 20 patients/day × 18 min × $0.05/min = $18/day — vs $11–$30/day for subscription scribes that cap usage.
3. Post-visit MEAT pass. A second LLM pass over the transcript, suspect list, and EHR context produces:
   • Suggested ICD-10 codes with specificity (E11.42 not E11.9)
   • MEAT-compliant assessment language for each chronic dx
   • Flagged conditions discussed but not assessed for clinician sign-off
4. Audit trail. Original audio + transcript + structured output retained per CMS RADV requirements (10-year retention if you bill Part C).

Vendor matrix — HCC-aware scribes 2026

For a panel under 1,000 MA lives, vendor pricing rarely beats DIY economics — especially if you already have a population-health analyst running gap-closure reports. For larger panels with prospective payment contracts, Navina-class tools pay for themselves on RAF lift alone within 90 days.

The DIY HCC scribe stack (under $200/provider/month)

If you're running a 1–3 provider clinic on full-risk MA contracts, this stack works in 2026:

• Audio capture: Otter Pro for Teams ($20/mo) or phone recorder + LessRec at $0.05/min for raw transcription
• EHR pull: Athena/eClinicalWorks/AdvancedMD APIs for problem list, labs, dx history (~$0 if your EHR exposes FHIR)
• HCC engine: Claude Sonnet or GPT-4-class via API, ~$0.20–$0.50 per encounter for the MEAT pass with the right prompt
• Suspect list builder: Python script over EHR data — flags conditions billed last year but not this year, conditions implied by labs but undocumented
• Compliance: Sign BAA with whichever LLM you use (Anthropic offers BAA, OpenAI Enterprise offers BAA), retain audio + transcript + structured output for 10 years

Total cost at 250 patients/month: ~$60–$140/month — vs $4,000–$10,000/month for vendor pricing on the same panel.

The MEAT prompt that catches what generic scribes miss

You are coding an MA encounter for HCC capture under CMS-HCC v28.

INPUT:
- Encounter transcript (verbatim)
- Patient problem list (last 24 months)
- Recent labs/imaging (last 12 months)
- Current medication list

For each CHRONIC condition discussed (or implied by data) in this encounter:
1. Cite the exact transcript line(s) that establish presence
2. Apply MEAT framework: which of Monitor / Evaluate / Assess / Treat is documented in the visit
3. Suggest ICD-10 code at maximum specificity supported by the encounter
4. Flag if specificity could be improved with one clinician question
5. Note if condition is being addressed but missing from formal assessment

Output ONLY conditions where MEAT is established in this encounter. Do not infer chronic conditions
that were merely mentioned without clinical engagement.

Flag separately: conditions implied by labs/imaging but never discussed in transcript — these are
clinician follow-up candidates, not codable for this encounter.

2026 audit risk: what RADV reviewers look for

CMS Risk Adjustment Data Validation (RADV) audits sample 30 charts per contract per year. Findings are extrapolated across the contract — one bad chart becomes 200 deductions. The 2026 audit focus areas:

• Status conditions claimed without supporting clinical activity (no MEAT)
• Specificity drops (E11.9 unspecified diabetes claimed at E11.42 specificity)
• "History of" used for active conditions to claim status without active management
• Mental health HCCs without DSM-5 criteria or duration documentation
• BMI-based HCCs (HCC 21) without documented complication or "morbid obesity" language

An AI scribe that outputs MEAT-compliant text and cites the transcript line for each code makes RADV defensible — the audit reviewer can trace dollar back to the patient's actual statement.

When to start

If your group bills Part C (Medicare Advantage) and your RAF score is below 1.05 with a panel of mostly seniors, you're almost certainly leaving 5–15% of capitation revenue on the table. Run a retrospective gap analysis on the last 6 months of MA encounters — expect 1.2–2.1 HCCs per patient that were in the EHR but not in the claim.

The DIY stack pays for itself in week one. Vendor stacks pay for themselves in month three on most panels. The only wrong answer is leaving v28 dollars unrecovered another year.