Value-based care depends on accurate diagnosis data. When information is lost between a clinical note and a risk score, budgets and quality benchmarks can fall out of line with the population being served. When a diagnosis is unsupported or overstated, it can create audit risk.
Two developments make this especially important for 2026. CMS materials move Medicare Advantage risk score calculations fully to the 2024 CMS-HCC model for calendar year 2026. ACO REACH also moves Standard and New Entrant ACOs to full use of the V28 prospective HCC model for performance year 2026. At the same time, Risk Adjustment Data Validation (RADV) activity continues to place greater scrutiny on the records behind submitted diagnoses.
AI can help organizations manage this work, but it does not change the underlying goals: capture what is clinically true, reduce unnecessary chart review, and maintain evidence for every reported code.
Key Takeaways
- Newer models now carry full weight. Medicare Advantage and ACO REACH organizations need workflows aligned with the models used for 2026.
- Audit readiness is an ongoing requirement. CMS uses RADV to identify unsupported diagnoses and estimate overpayments.
- Documentation determines the outcome. Diagnoses used for payment must be supported by the medical record. MEAT, meaning monitoring, evaluating, assessing, or treating, remains a practical documentation guide.
- AI works best with human review. Models can screen records and prioritize evidence, while coders and clinicians make final decisions.
- Deletions matter as much as additions. A defensible workflow removes unsupported codes instead of focusing only on missed conditions.
- Results should be measured narrowly. Track evidence-backed acceptance rates, documentation completeness, and audit turnaround rather than relying on projected revenue claims.
Risk Adjustment Basics for Value-Based Care
Risk adjustment translates a patient population’s clinical complexity into a payment and benchmarking signal. Hierarchical Condition Categories (HCCs) group ICD-10-CM diagnoses into payment-related categories. Those categories are combined with demographic factors to produce a risk adjustment factor, or RAF, for each member.
Two features shape daily operations. First, risk scores generally reset each year, so a chronic condition documented last year does not automatically carry forward. Second, a code is only as reliable as the note supporting it. The American Academy of Family Physicians describes MEAT as a useful rule of thumb for deciding whether documentation supports a reported condition.
Program context also matters. Medicare Advantage contracts are subject to CMS RADV, while the Affordable Care Act individual and small-group markets use a separate process commonly called HHS-RADV. The mechanics differ, but both require the submitted record to support the diagnosis without added explanation from the organization.
What Changes for 2026
- Audit sampling has a defined range. For payment year 2020 and later audits, CMS may sample 35 to 200 enrollees per Medicare Advantage contract to estimate overpayments.
- Process guidance is more centralized. CMS has consolidated RADV program information, including material on intake and abstraction reviews performed by certified medical record coders.
- Audit cycles require active monitoring. Organizations should confirm current payment-year schedules and submission deadlines through the latest CMS notices.
- Model updates place greater weight on accuracy. The 2024 CMS-HCC model uses ICD-10-based categories intended to improve prediction and reduce the payment effects of coding variation.
Where AI Helps Across the Lifecycle
Prospective review. Before a visit, natural language processing can scan prior notes, laboratory results, pharmacy data, and claims for conditions that appear clinically active but have not been documented this year. The system should provide a short, ranked list with supporting text rather than an unfiltered set of codes. A prompt in the electronic health record can then let the clinician confirm, refine, or dismiss each suggestion.
Concurrent review. After an encounter, AI can compare the documentation with the submitted codes. This helps teams address gaps while the visit is still recent, reducing the need for retrospective clinician queries months later.
Retrospective review. At larger volumes, AI can help identify justified additions and flag codes that the record does not support. Vendor approaches vary. Cotiviti, for example, describes AI-assisted pre-visit preparation and post-visit HCC reconciliation. RAAPID describes a neuro-symbolic approach, combining language analysis with defined rules, that connects suggested HCCs with MEAT evidence. These are vendor-stated capabilities and should be validated in the organization’s own environment.
Build Audit-Defensible Outputs
Regulatory expectations should translate into clear system requirements. A defensible workflow links every HCC to the relevant source text and date of service. It also records who reviewed each suggestion, what decision was made, and when the review occurred. Teams should be able to retrieve documentation for an audit sample and export it in the required format.
Mature programs apply the same standard to additions and deletions. A process focused only on finding new HCCs introduces one-directional bias. Routing unsupported codes for removal with equal care produces cleaner data and reduces the number of discrepancies that must be explained during an audit.
A Practical Implementation Plan
Days 1 to 30. Establish data access and governance for EHR, claims, laboratory, and pharmacy information. Validate V28 and 2024 CMS-HCC mappings against a known sample. Measure current documentation completeness and chart-request volume to create a baseline.
Days 31 to 60. Pilot the workflow with a limited set of high-variance conditions rather than the entire HCC library. Create a coder quality-assurance queue with clear acceptance criteria. Give clinicians short guidance that explains MEAT in clinical terms.
Days 61 to 90. Run a mock RADV exercise from sample intake through preparation of the submission package. Measure the time required and review disagreements between AI suggestions and coder decisions. Use those findings to adjust prompts, evidence thresholds, or mappings.
Train clinicians on documentation sufficiency and provide feedback at the encounter level. Do not rely on a health risk assessment as the only evidence source, and do not treat a model suggestion as a final coding decision.
What to Look for in a Platform
Start with practical requirements: integration with the existing EHR, visible evidence for every suggestion, prospective and retrospective review, a RADV workspace for sample intake and export, coder quality controls, role-based approval, documented HIPAA and SOC 2 practices, and suitable deployment options through an API or software-as-a-service model.
Platforms that support both prospective gap review and retrospective audit preparation may reduce duplicated work because both workflows depend on the same evidence. One commercial risk adjustment solution, RAAPID, combines point-of-care prompts, HCC evidence review, and RADV workflow functions. Treat the RAAPID description, like any vendor description, as a starting point for a controlled pilot rather than proof of performance.
Measure Impact Without Overclaiming
Useful measures are narrow enough to verify internally. Examples include the percentage of AI suggestions accepted with documented MEAT support, the balance between confirmed additions and unsupported deletions, chart-request volume, audit-package turnaround time, documentation completeness, and performance against sample-to-submission deadlines.
Avoid presenting results as guaranteed revenue growth or protection from audits. Software cannot control either outcome. Results also depend on clinician documentation, coder review, data quality, and the organization’s governance practices.
Governance and Guardrails
Set the policy before expanding a pilot. Require a qualified coder to approve every coding change, route ambiguous cases to a clinician, retain complete audit logs, and maintain a manual process for periods when the model is unavailable. Review samples regularly for drift, especially after model or mapping updates.
Align internal standards with AHIMA ethical coding guidance and AAFP documentation guidance so outside reviewers can recognize the methods being used. Whether an organization builds its own workflow or evaluates a platform such as RAAPID, governance is what makes the technology defensible rather than simply faster.
The 2026 Priority
Model changes, active audit programs, and more capable AI tools are arriving at the same time. CMS has also described using AI to screen records before human RADV review, reinforcing the value of combining automated analysis with professional judgment.
Success should not be measured by the number of suggestions a system generates. It should be measured by whether documentation stands on its own, coders can verify the evidence efficiently, and clinicians receive useful prompts without unnecessary disruption. When those elements are in place, AI-supported risk adjustment can improve data accuracy while helping value-based care programs align resources with documented patient needs.
