How AI Scribes Document Medicare Advantage V28 HCC Recapture for 2026

AI scribes document Medicare Advantage V28 HCC recapture for 2026 by capturing each chronic condition during a face-to-face or video encounter, linking it to explicit MEAT evidence (Monitor, Evaluate, Assess, Treat), coding to V28-valid ICD-10-CM specificity, and requiring clinician attestation before any diagnosis reaches the claim.

This playbook is for Clinic Medical Directors who own risk adjustment accuracy, audit exposure, and clinician burden at the same time. It covers what changed when V28 reached full weight, which diagnoses lost payment relevance, how ambient documentation must be structured to survive Risk Adjustment Data Validation (RADV), and where AI assistance ends and clinician judgment begins.

V28 reaches full weight in 2026, ending V24 blended revenue

CLINICAL UPDATE 2026: Revised for 2026 CMS CPT G2211 attestation rules, California SB 1120 compliance, and NCCI Modifier 25 bundling criteria.

CMS phased in the 2024 CMS-HCC model over three payment years: 33 percent V28 in 2024, 67 percent in 2025, and 100 percent in 2026. For payment year 2026, no V24 blending remains to cushion diagnoses that lost value. Every risk score your patients carry is now calculated entirely under V28 logic.

The structural changes are significant. V28 expanded to 115 payment HCCs (from 86 in V24), reorganized hierarchies, and removed payment relevance from roughly 2,200 ICD-10-CM codes. CMS also applied constraints that assign identical coefficients to clinically related HCCs, most notably across the diabetes family, so complication documentation now drives accuracy differently than it did under V24.

The practical consequence for clinics: recapture programs built on V24 suspect lists now generate prompts for conditions that no longer move the risk score, while missing the specificity V28 actually rewards. An AI scribe trained or configured on V24 logic actively wastes clinician attention in 2026.

HCC recapture is a documentation event, not a coding event

Risk adjustment resets every January. A diagnosis documented in 2025 does not carry into 2027 payment unless a qualifying provider reassesses and documents it during a 2026 date of service. "Recapture" is the annual act of re-documenting a persistent chronic condition with current clinical evidence.

Qualifying encounters have strict boundaries. CMS accepts diagnoses from face-to-face encounters with acceptable provider types (physicians, nurse practitioners, physician assistants, and other recognized clinicians) and from telehealth visits that use real-time audio and video. Audio-only encounters do not qualify for risk adjustment submission. Your AI scribe must record encounter modality as structured data, not buried narrative.

Coders cannot rescue thin notes. If the clinician mentions "diabetes, stable" with no evaluation, no data review, and no plan, no coder can legitimately assign a complication code. The note itself must carry the clinical reasoning. This is why recapture success in 2026 depends on how the AI scribe structures the assessment and plan, not on downstream coding software.

MEAT evidence must live inside the encounter note for every HCC

MEAT is the auditor's lens. RADV reviewers and plan compliance teams look for at least one MEAT element tied to each submitted diagnosis within the same date of service. Best practice for 2026 is two or more elements per HCC, explicitly linked to the diagnosis rather than scattered across the note.

• Monitor: signs, symptoms, progression. "Reports two-pillow orthopnea unchanged since last visit; weight stable at 184 lb."

• Evaluate: test results, exam findings. "eGFR 41 on 1/14/2026, down from 47; UACR 210 mg/g."

• Assess: status and clinical judgment. "Chronic diastolic heart failure, compensated, NYHA class II."

• Treat: medications, referrals, orders. "Continue empagliflozin 10 mg daily; refer to nephrology."

Problem lists are not evidence. A diagnosis appearing only in the problem list, past medical history, or medication reconciliation section is generally insufficient support. Auditors expect the assessment and plan to address the condition. AI scribes that auto-populate the assessment from the problem list create the exact pattern OIG has repeatedly flagged.

Rule: No MEAT in the assessment, no HCC on the claim.

V28 specificity rules change what AI scribes should prompt clinicians to clarify

Effective recapture prompting in 2026 means asking clinicians to clarify specificity that V28 rewards, while dropping prompts for codes that no longer carry payment weight. The following conditions account for most recapture volume in adult primary care and cardiology panels.

Diabetes complications require documented causal linkage

V28 constrains diabetes HCCs so that diabetes with and without complications carry equal coefficients. That reduces the financial incentive to code complications, but it does not reduce the accuracy requirement. Documenting E11.22 (ICD-10-CM) (type 2 diabetes with diabetic chronic kidney disease) still matters because it requires the companion N18 stage code, and the CKD stage maps to its own HCC.

The note must connect them. ICD-10-CM guidelines presume a link between diabetes and CKD unless the clinician states otherwise, but auditors still expect the assessment to address both conditions with current data. The scribe should capture the eGFR trend, UACR, and the plan (SGLT2 inhibitor, ACE inhibitor or ARB, nephrology referral) under a combined assessment heading.

CKD stage 3 now splits into 3a and 3b

V28 separates CKD stage 3B (N18.32) from stage 3 unspecified and 3A (N18.30, N18.31) into distinct HCCs. Unspecified stage 3 underrepresents a stage 3B patient. AI scribes should prompt when lab values support 3B (eGFR 30 to 44) but the clinician says only "CKD 3."

Heart failure type and acuity drive HCC assignment

Heart failure in V28 spans several HCCs differentiated by acuity and severity, including end-stage, acute on chronic, acute, and heart failure except end-stage and acute. I50.32 (ICD-10-CM) (chronic diastolic heart failure) maps to the chronic category, but only when type and chronicity are documented. "CHF, stable" defaults to I50.9, which loses clinical precision and invites query.

Ambient capture helps here. Clinicians frequently state ejection fraction, volume status, and diuretic adjustments aloud during the visit. A well-designed scribe converts that conversation into a structured assessment: type (HFpEF or HFrEF), chronicity, current status, and treatment changes.

Several common diagnoses lost payment relevance entirely

V28 removed payment weight from mild and unspecified major depressive disorder, stable angina pectoris, atherosclerosis of native extremity arteries without ulceration or gangrene, and mild or moderate protein-calorie malnutrition. Clinicians must still document these conditions accurately for care quality, but AI scribes should stop flagging them as recapture priorities.

Severity documentation now matters more. Moderate or severe major depression retains HCC value, so the PHQ-9 score and clinician severity statement belong in the assessment. For peripheral artery disease, documenting ulceration or gangrene when present (with laterality and site) is essential, but the scribe must never suggest a complication the exam does not support.

Each V28 clinical parameter maps to a distinct audit vulnerability

Every row shares one principle: the AI drafts and proposes, while the clinician decides. Specialty-specific versions of these controls are available in the Specialty Clinical Documentation Library, and prebuilt assessment structures for each HCC family live in 1,000+ Specialty Clinical Macros.

Longitudinal reasoning, not simple retrieval, prevents carried-forward diagnosis risk

Recapture is inherently longitudinal. The scribe must know that heart failure was assessed in March 2025, that the patient has not been seen since, and that today's visit is the first 2026 opportunity to reassess. It must also recognize when a condition has resolved, such as a cancer now in remission beyond active treatment, and stop prompting for it.

Retrieval-based systems struggle here. Many AI scribes retrieve fragments of prior notes and paste them into today's draft. That approach reproduces stale diagnoses as if they were current findings, which is precisely the cloned documentation pattern auditors look for. The failure modes are explained in Why RAG-Based AI Scribes Fail Longitudinal Reasoning.

The correct architecture separates context from evidence. Prior history informs what the scribe asks the clinician to address, but only content spoken, examined, reviewed, or ordered during today's encounter becomes MEAT evidence. Medical Directors should test any vendor by asking one question: can the system show, for each proposed HCC, the exact transcript or data source from today's visit that supports it?

Test: Every proposed HCC needs a same-day evidence source.

G2211 and Modifier 25 require separate attestation from HCC recapture

Recapture visits often overlap with annual wellness visits and longitudinal care billing, which creates bundling risk if documentation blurs the services. HCC recapture itself is not billable; it rides on the E/M or AWV encounter. Your billing logic and your documentation logic must stay aligned.

G2211 reflects ongoing longitudinal care. The add-on code applies when the practitioner serves as the continuing focal point for a patient's care or provides ongoing care for a single serious or complex condition. Since CY 2025, Traditional Medicare permits G2211 with a Modifier 25 E/M when the other same-day service is an AWV, vaccine administration, or a Part B preventive service. Medicare Advantage plans may apply their own payment policies, so verify contract terms.

Modifier 25 demands separation. Under NCCI principles, a problem-oriented E/M billed with an AWV must be significant and separately identifiable. If the clinician adjusts heart failure medications during an AWV, the scribe should structure two distinct sections: the AWV components (health risk assessment, prevention plan) and the problem-oriented assessment with its own MEAT evidence and medical decision making.

The attestation should be explicit. A brief clinician statement such as "I am the continuing focal point for this patient's chronic heart failure and diabetes care" supports G2211. A separate statement identifying the problem-oriented work supports Modifier 25. The AI scribe can draft both, but the clinician must confirm both apply.

California SB 1120 reinforces physician judgment over AI-assisted determinations

California SB 1120 took effect January 1, 2025, and requires that health plans and disability insurers using AI in utilization review leave medical necessity determinations to a licensed physician or qualified health care provider. Although the statute targets payer utilization management, it signals the regulatory direction for all clinical AI: algorithms support decisions; licensed clinicians make them.

Apply the principle internally. Clinics operating in California, or contracting with California-regulated plans, should document that every diagnosis submitted for risk adjustment reflects clinician judgment. The safest posture is a diagnosis-level accept or reject action captured in the audit trail, with the AI's proposed text and the clinician's final text both preserved.

Other states are following similar paths, and CMS has signaled heightened scrutiny of AI-driven decisions in Medicare Advantage. Building human-in-the-loop controls now avoids retrofitting workflows later.

Medical Directors can run 2026 recapture in six governed steps

1. Refresh your suspect logic. Remove V24-only targets and rebuild prompts around V28 HCCs, constrained families, and new specificity splits such as CKD 3B.

2. Prepare charts before visits. Surface chronic conditions not yet assessed in 2026 as pre-visit context for the clinician, labeled clearly as historical, never as current findings.

3. Capture the encounter ambiently. Let the scribe record the conversation, exam narration, and data review, then structure each discussed condition into a MEAT-linked assessment block.

4. Query specificity gaps in real time. Prompt the clinician only when clinical evidence suggests greater specificity (stage, type, severity, complication) and only for conditions actually addressed.

5. Enforce diagnosis-level attestation. Require the clinician to accept, edit, or reject every proposed diagnosis before signature, and log each action.

6. Audit a monthly sample. Pull ten notes per clinician, verify MEAT linkage for each HCC, and track rejection rates by HCC family to tune prompts.

Rejection rates are diagnostic data. If clinicians reject a high share of proposed diabetes complication codes, your prompts are over-suggesting. If they rarely reject anything, verify that review is genuine rather than reflexive clicking, since unreviewed acceptance offers no audit protection. The Scribing.io Clinical Interface displays these acceptance metrics by clinician and HCC family.

RADV extrapolation makes clinician attestation the final control point

The RADV landscape changed sharply. The 2023 RADV final rule authorized extrapolation of error rates beginning with payment year 2018, and in 2025 CMS announced plans to audit all eligible MA contracts each year with expanded staffing and technology. Errors are no longer isolated chart findings; they become contract-wide repayment multipliers.

Plans are pushing risk downstream. Expect your MA contracts to include stronger documentation warranties, retrospective review clauses, and recoupment provisions tied to unsupported HCCs. A provider group whose notes show consistent MEAT linkage becomes a lower-risk partner in value-based negotiations.

OIG and DOJ scrutiny continues. OIG has repeatedly highlighted diagnoses supported only by health risk assessments or chart reviews with no follow-up care, and False Claims Act settlements in Medicare Advantage have reached substantial figures. AI tools that inflate diagnosis capture without clinical support amplify this exposure at scale.

Accuracy, not volume, is the goal. The best 2026 recapture program produces fewer total HCC submissions than an aggressive suspecting program but with near-complete support on audit. Medical Directors should report recapture rate and audit support rate together, never one without the other.

Metric pair: Track recapture rate alongside audit support rate.

Frequently asked questions about Medicare Advantage V28 HCC recapture

Does V28 require annual recapture of every chronic condition?

Yes. Risk adjustment resets every January, so each chronic condition must be documented during a qualifying face-to-face or audio-video encounter in the calendar year with MEAT support. A diagnosis captured in 2025 contributes nothing to 2027 payment unless clinicians reassess and document it again during 2026.

Which ICD-10-CM codes lost payment value under V28?

V28 removed payment value from mild and unspecified major depression, stable angina, atherosclerosis of native extremity arteries without ulceration or gangrene, and mild or moderate protein-calorie malnutrition, among roughly 2,200 codes. Clinicians still document these conditions accurately, but specificity such as ulceration, severity, or exacerbation now determines HCC eligibility.

Can an AI scribe add a diagnosis the clinician did not discuss?

No. An AI scribe may surface prior diagnoses as prompts for clinician review, but it must never insert an assessed diagnosis absent from the encounter conversation. Unsupported diagnoses create False Claims Act exposure, and RADV auditors will not accept unattested AI-generated problem lists as valid support.

Can G2211 be billed with Modifier 25 on an AWV day?

Yes, since CY 2025, CMS permits G2211 with a Modifier 25 E/M when the same-day service is an annual wellness visit, vaccine administration, or Medicare Part B preventive service. Documentation must still show the longitudinal relationship and a significant, separately identifiable problem-oriented service distinct from the AWV.

How does RADV extrapolation affect AI-documented HCCs?

CMS announced in 2025 that it will audit every eligible MA contract annually and extrapolate error rates across the full contract. One unsupported HCC pattern, such as carried-forward heart failure without MEAT, can multiply into large repayments. Encounter-level MEAT linkage and clinician attestation remain the primary defenses.