Improving HCC Risk Adjustment with AI: The 2026 Clinical Library Playbook for V28 RAF Recapture

Clinical Update — June 2026: This guide has been revised to reflect the final CMS-HCC V28 phase-in weights effective for 2026 DOS (67% V28 / 33% V24) and the 2027 transition to 100% V28 weighting. Updated sections include KDIGO 2024 staging alignment, FHIR R4 Encounter.diagnosis binding specifications, and the 2026 HIPAA patient consent requirements for ambient AI scribes. If you implemented workflows based on the prior version, review H2 sections 2 and 5 for material changes.

TL;DR

Under the 2026 CMS-HCC V28 model, unspecified chronic condition codes (N18.9 for CKD, E11.9 for T2DM without complications) no longer reliably risk-adjust, threatening Medicare Advantage capitation payments. This playbook details how AI-driven ambient scribes—specifically Scribing.io—solve the RAF recapture problem by auto-staging conditions from structured lab data, generating compliant MEAT documentation in real time, and exporting FHIR R4 audit packets with full lab provenance. The result: compliant, auditable HCC capture without upcoding, and measurable protection against 2027 capitation drops. If you manage risk adjustment or CDI for a Medicare Advantage organization, this is the operational blueprint your competitors don't have.

• Why V28 Changes Everything for HCC Risk Adjustment in 2026

• The Information Gain Competitors Miss — Dual-Code Staging, Lab Provenance, and FHIR Auditability

• Scribing.io Clinical Logic — Handling the Q4 2026 Annual Wellness Visit Scenario

• Technical Reference: ICD-10 Documentation Standards

• Regulatory Compliance Architecture: HIPAA, RADV, and State AI Laws

• Implementation Operations Playbook: 30/60/90-Day Deployment

• Measuring RAF Recapture ROI Under V28

Why V28 Changes Everything for HCC Risk Adjustment in 2026

The CMS-HCC V28 risk adjustment model, now in its final phase-in year, represents the most consequential shift in Medicare Advantage payment methodology since the original HCC framework launched. For Directors of Risk Adjustment and Clinical Documentation Integrity, the operational reality is blunt: codes that captured RAF value under V24 have been reclassified, merged, or eliminated. Conditions that once carried standalone HCC mappings now demand clinical specificity that most EHR workflows were never designed to produce.

Scribing.io exists to close this gap—not through retrospective chart chases or query-based CDI, but through real-time ambient documentation that captures condition specificity, stages from structured lab data, and exports audit-ready FHIR R4 resources at the point of care. Before detailing how, let's establish exactly what V28 demands and why legacy workflows fail.

What V28 Specifically Demands

The CMS V28 final rule introduced structural changes that directly erode HCC capture rates for organizations still running V24-era documentation workflows:

• Condition specificity thresholds increased. Unspecified codes (e.g., N18.9 for CKD, unspecified) map to fewer or no HCCs. E11.9 (T2DM without complications) no longer captures the RAF value that E11.22 (T2DM with diabetic CKD) does. The hierarchical reassignment means organizations lose credit for conditions they clinically manage but fail to document at full specificity.

• Dual-code requirements for comorbid conditions. Diabetic kidney disease requires BOTH the diabetes-with-CKD combination code AND a separate CKD stage code. Omitting either one creates an incomplete clinical picture that may fail risk adjustment entirely. This dual-code architecture is not new to ICD-10 conventions, but V28 penalizes its absence with HCC non-mapping for the first time at this scale.

• Annual recapture mandates remain absolute. Every chronic HCC must be documented with clinical evidence every calendar year. A condition coded in 2025 but not re-documented with MEAT elements in 2026 drops from the 2027 RAF calculation. There is no carryover, no grace period, no exception.

• MEAT documentation is the audit standard. Monitor, Evaluate, Assess, Treat—all four elements must appear in the encounter note for the condition to be considered "addressed" per CMS RADV audit standards. A problem list entry alone does not suffice. A mention in the HPI without treatment linkage does not suffice.

The Revenue Impact Is Not Hypothetical

Organizations failing to recapture chronic HCCs under V28 face RAF score reductions of 5–15% per affected member. For a mid-sized MA plan with 40,000 members, even a 3% aggregate RAF erosion translates to $8–12 million in annual capitation loss, depending on the member mix's clinical acuity. The 2026 CMS Rate Announcement confirmed the 67% V28 / 33% V24 blended weights for payment year 2026, but 2027 moves to 100% V28—making this calendar year the last operational window to build compliant recapture workflows before the full financial impact materializes.

The AMA/AHIP/NAACOS Value-Based Care Best Practices Playbook (2024) acknowledges risk adjustment as one of seven core VBC payment domains. It discusses the importance of data completeness and multi-payer alignment. What it does not provide—and what this playbook addresses—is the clinical decision logic, FHIR-level technical architecture, and real-time documentation workflow required to actually execute compliant HCC recapture at scale under V28.

The Information Gain Competitors Miss — Dual-Code Staging, Lab Provenance, and FHIR Auditability

The most significant gap in existing risk adjustment guidance—and in competing AI documentation platforms—is the failure to address condition-level coding architecture under V28. Specifically: the dual-code requirement for diabetic kidney disease and the necessity of lab-sourced staging provenance that survives RADV audit.

The Dual-Code Problem

In 2026 CMS-HCC V28, compliant RAF recapture for diabetic kidney disease requires BOTH the diabetes-with-CKD code (E11.22 - Type 2 diabetes mellitus with diabetic chronic kidney disease; N18.4 - Chronic kidney disease) AND a separate CKD stage code (e.g., stage 4 (severe)). Unspecified staging—coding N18.9 instead of N18.4—often fails to risk-adjust under V28 because the HCC mapping requires stage specificity to assign the condition to the correct hierarchical category.

This is not a theoretical compliance concern. It is the primary mechanism by which RAF value evaporates at the encounter level. A coder working from a note that says "CKD—stable" has no clinical basis to assign a staged code. The coder correctly selects N18.9. The organization loses the HCC. The capitation payment drops. No one committed fraud. The documentation simply lacked specificity.

Compliance with California SB-1120 and other state AI utilization review laws further constrains how AI systems can interact with clinical decision-making—meaning the solution must preserve physician authority while still closing documentation gaps. This is a design constraint most competitors have not solved.

Why NLP-Only Approaches Fail

Most AI scribe and CDI platforms focus on NLP extraction from unstructured text—scanning the narrative for mentions of conditions and suggesting codes. This approach fails the V28 dual-code problem for three specific reasons:

1. The staging data isn't in the narrative. CKD stage is determined by laboratory values (eGFR and UACR), not by clinician dictation. If the clinician says "CKD—stable," there is no text to extract a stage from. The KDIGO 2024 CKD staging criteria define stage 4 as eGFR 15–29 mL/min/1.73m². That data lives in the lab interface, not the progress note.

2. The treatment linkage isn't explicit. V28 compliance requires that the Assessment & Plan connect the staged condition to active treatment decisions. Generic CDI tools flag missing diagnoses but don't generate the treatment-condition binding that RADV auditors require.

3. There is no audit provenance. Even when a code is correctly assigned, payers conducting RADV audits require traceability from code → note → clinical evidence → objective data. Without structured lab references in the encounter data, the audit trail breaks at the evidentiary link.

How Scribing.io Solves This: Three-Layer Architecture

Layer 1: Structured Lab Ingestion. The system reads structured lab data via standard LOINC codes—specifically LOINC 33914-3 (eGFR by creatinine-based formula) and LOINC 9318-7 (Albumin/Creatinine ratio in urine). Values are date-stamped and sourced from the EHR's lab interface via FHIR R4 Observation resources, not from clinician dictation. This produces a staging determination with objective, auditable provenance per NIH-published KDIGO staging criteria.

Layer 2: Real-Time Micro-Prompting. When the system detects that a condition's stage or treatment linkage hasn't been verbally affirmed by the provider, it issues a contextual micro-prompt—e.g., "Confirm CKD stage based on eGFR 24?" This preserves clinician authority over the clinical decision while ensuring the documentation captures the specificity required for coding. The provider retains full control; the system surfaces data, not diagnoses.

Layer 3: FHIR R4 Audit Export. The system writes auditable FHIR R4 data structures: Encounter.diagnosis references a coded Condition resource (E11.22) plus a staged CKD Condition with Condition.stage.summary and Evidence.detail → Observation, linking directly to the date-stamped lab values. Payers conducting RADV audits trace from the HCC code through the encounter, to the condition resource, to the lab observation, with unbroken provenance.

Scribing.io Clinical Logic — Handling the Q4 2026 Annual Wellness Visit Scenario

This section presents the centerpiece clinical scenario demonstrating Scribing.io's operational value for HCC risk adjustment under V28. This is the difference between risk adjustment guidance and risk adjustment execution.

The Scenario

Q4 2026. A primary care physician sees a 74-year-old Medicare Advantage patient with Type 2 diabetes mellitus and chronic kidney disease during an Annual Wellness Visit. The physician's standard AWV template produces an assessment of "CKD—stable" without stage specification, without lab value references, and without treatment-condition linkage.

Without Scribing.io: The Default Failure Mode

The coder receives the note. "CKD—stable" provides no clinical basis for stage assignment. Per AMA ICD-10-CM coding guidelines, the coder correctly assigns:

• E11.9 — Type 2 diabetes mellitus without complications

• N18.9 — Chronic kidney disease, unspecified

Under V28, E11.9 captures minimal RAF value because the diabetes-CKD causal relationship is not documented. N18.9 maps to no HCC because CKD stage is unspecified. The patient's RAF score drops. Multiply this across a panel of several hundred similar patients and the organization faces a measurable capitation reduction in 2027.

No compliance violation occurred. No one upcoded. The documentation simply failed to represent the clinical reality that the provider knows and manages.

With Scribing.io: Step-by-Step Clinical Logic

The ambient AI scribe runs during the encounter. The Anchor Truth governing every system action: Under the 2026 CMS V28 Model, AI must automate the MEAT (Monitor, Evaluate, Assess, Treat) documentation for every chronic condition annually to ensure accurate RAF Recapture and prevent capitation payment drops.

The Outcome

The RAF score accurately reflects the patient's clinical complexity. The 2027 capitation payment is protected. The documentation withstands RADV audit because every code traces to objective clinical evidence through the FHIR resource chain. The clinician spent approximately 8 additional seconds confirming the CKD stage. No upcoding occurred—the stage assignment is provably derived from lab data within KDIGO staging criteria and confirmed by the treating physician.

See it on your own charts: Scribing.io offers 2026 CMS-HCC V28 MEAT auto-generation with lab-driven CKD staging, FHIR Encounter.diagnosis binding, and a 1-click RADV audit packet—run the demo with your data.

Technical Reference: ICD-10 Documentation Standards

This section details the ICD-10-CM coding specificity requirements that Scribing.io enforces in real time to prevent denials and ensure V28 HCC mapping.

The Specificity Mandate

Under the CMS ICD-10-CM Official Guidelines, coders are required to assign codes to the highest level of specificity supported by the clinical documentation. When documentation lacks specificity, the coder must assign an unspecified code—even when the clinical team manages the condition at a higher level of detail. This creates the documentation-coding gap that V28 punishes.

Diabetic Kidney Disease: The Dual-Code Requirement

For diabetic kidney disease, ICD-10-CM conventions require two codes:

1. The combination code establishing the causal relationship: E11.22 - Type 2 diabetes mellitus with diabetic chronic kidney disease; N18.4 - Chronic kidney disease. This code establishes that the diabetes and kidney disease are clinically linked—a determination that requires explicit provider documentation per AMA coding guidelines. The provider must document that the CKD is attributable to the diabetes; a co-occurring mention alone is insufficient.

2. The staging code specifying CKD severity: stage 4 (severe). N18.4 maps to HCC 329 (Chronic Kidney Disease, Stage 4) under V28, which carries significant RAF weight. N18.9 (unspecified) does not reliably map to this HCC.

How Scribing.io Ensures Maximum Specificity

The JAMA Network Open analysis of HCC coding accuracy found that documentation specificity gaps account for a larger share of RAF score erosion than outright missed diagnoses. The clinical condition is known. The documentation simply doesn't say it at the specificity level the code set requires. Scribing.io targets this exact failure mode.

Regulatory Compliance Architecture: HIPAA, RADV, and State AI Laws

Deploying AI for risk adjustment documentation triggers overlapping regulatory requirements. Scribing.io's compliance architecture addresses three regulatory layers simultaneously.

HIPAA 2026 and Ambient AI Consent

The 2026 HIPAA update on patient consent for ambient AI scribes established new requirements for disclosure and authorization when ambient AI systems process patient encounters. Scribing.io implements:

• Pre-encounter disclosure templates integrated into the check-in workflow, documenting patient awareness that an AI scribe will process the encounter audio

• BAA-compliant processing with all PHI handled under executed Business Associate Agreements between the practice, Scribing.io, and any sub-processors

• Audio retention policies configurable per practice, with default auto-deletion after note finalization and attestation, per the minimum necessary standard

RADV Audit Preparedness

CMS RADV audits require that every risk-adjusted condition code be supportable by encounter-level clinical documentation. The audit standard is not "was the code reasonable?" but "can the auditor trace the code to clinical evidence in the medical record?" Scribing.io's FHIR R4 audit packet provides:

• The encounter note with MEAT documentation

• The FHIR Condition resource with code, stage.summary, and evidence.detail

• The linked Observation resources with lab values, LOINC codes, and collection dates

• The MedicationRequest and ServiceRequest (referral) resources linked to the condition

• The clinician attestation timestamp

This bundle is exportable as a single JSON document or rendered PDF, reducing RADV response preparation from hours per chart to under 2 minutes.

State AI Laws

Fourteen states had enacted or proposed AI-in-healthcare legislation as of Q2 2026. California SB-1120 is the most prescriptive, requiring that AI systems used in utilization review or clinical decision support disclose their role and preserve human override authority. Scribing.io's micro-prompt architecture inherently satisfies this requirement: the system surfaces lab-derived staging data and requests provider confirmation. It does not assign codes autonomously. The clinician confirms, modifies, or rejects every suggestion. This is documented in the attestation record.

Implementation Operations Playbook: 30/60/90-Day Deployment

For Directors of Risk Adjustment and CDI deploying Scribing.io across a Medicare Advantage provider network, this is the operational timeline.

Critical Implementation Note

The 30-day technical integration phase is not optional or compressible. Lab feed validation—confirming that the EHR's FHIR Observation resources include accurate LOINC codes, reference ranges, and collection timestamps—is the foundation of the entire staging and MEAT architecture. Organizations that skip this step will generate staging suggestions based on incomplete or mismatched lab data, creating compliance risk rather than reducing it.

Measuring RAF Recapture ROI Under V28

RAF recapture ROI is measurable at the encounter level, the provider level, and the plan level. The following framework maps directly to the metrics that MA plan CFOs and risk adjustment directors report to CMS and to their boards.

Encounter-Level Metrics

• HCC recapture rate: Percentage of chronic conditions with active HCC mappings that are successfully re-documented with compliant MEAT in the current calendar year. Target: ≥90% for conditions with prior-year RAF value.

• Specificity upgrade rate: Percentage of encounters where the AI micro-prompt resulted in a code specificity upgrade (e.g., N18.9 → N18.4). This metric directly quantifies documentation-driven RAF recovery.

• Dual-code compliance rate: For conditions requiring combination codes (E11.22 + N18.4), the percentage of encounters where both codes are assigned with supporting documentation. Target: ≥95%.

Provider-Level Metrics

• Micro-prompt acceptance rate: Percentage of system prompts where the provider confirmed or modified the suggestion (vs. dismissed without response). Consistently low rates (<70%) for a specific provider indicate training needs or workflow friction.

• MEAT completeness score: Automated assessment of whether all four MEAT elements are present for each chronic condition addressed in the encounter. Scored per condition per visit.

Plan-Level Financial Metrics

• RAF score variance, year-over-year: Comparison of aggregate RAF scores for the same member cohort before and after Scribing.io deployment. This isolates documentation-driven RAF changes from changes attributable to member health status.

• Projected capitation recovery: Calculated as: (RAF score increase per member) × (CMS base rate) × (member months). For a plan with 40,000 members and a 0.03 average RAF score recovery, the annual capitation recovery at a $1,100 base rate ≈ $15.8M.

• RADV audit performance: Percentage of audited charts where the documentation fully supports the coded conditions without amendment. Target: ≥97% with Scribing.io audit packets.

The Bottom Line

Risk adjustment under V28 is a documentation execution problem, not a clinical knowledge problem. Providers know their patients have stage 4 CKD. They manage it with ACE inhibitors and SGLT2 inhibitors and nephrology referrals. The medical record just doesn't say so at the specificity level that V28 demands, with the MEAT structure that RADV auditors require, linked to the lab evidence that proves the staging is objective.

Scribing.io closes this gap in real time, at the point of care, with 8 seconds of provider effort per condition. The codes are compliant. The documentation is auditable. The RAF is recaptured.

Run it on your own charts: Request the Scribing.io demo and see 2026 CMS-HCC V28 MEAT auto-generation with lab-driven CKD staging, FHIR Encounter.diagnosis binding, and a 1-click RADV audit packet—built for the way risk adjustment actually works in 2026.