HCC V28 Risk Adjustment: How Agentic AI Closes the Gap

HCC V28 Risk Adjustment: How Agentic AI Closes the Gap

As of January 1, 2026, CMS-HCC Model V28 is fully operative for Medicare Advantage risk scoring -- no more phased-in blend with V24. The transition dropped the number of valid ICD-10-CM diagnostic codes from 9,797 to 7,770, expanded hierarchical condition categories from 86 to 115, and is projected to reduce average Medicare Advantage risk scores by 3.12 percent across the board. Then, on April 1, 2026, CMS added 487 new ICD-10-CM codes -- most concentrated in injury, poisoning, and skin disorders -- reshuffling which diagnoses map to which HCCs. For medical coders and CDI specialists working in Medicare Advantage, the combination is a revenue problem that demands a systematic answer. Agentic AI is becoming that answer.

What V28 Actually Changed -- and Why It Hurts

V24 was forgiving with unspecified codes. A diagnosis of chronic kidney disease, unspecified (N18.9) still triggered HCC 136 and produced a risk adjustment factor. V28 eliminates most of that flexibility. Under V28, kidney disease without stage specificity earns no HCC, no RAF, and no payment. The same pattern repeats across diabetes complications, heart failure subtypes, and several neurological conditions. Coders who were trained on V24 logic and haven't retooled face a silent revenue bleed: the claim processes, the encounter looks clean, but the risk score never materializes.

The April 2026 ICD-10-CM additions compound this. New combination codes replace old two-code sequences in several chronic disease categories. A condition previously coded with a primary diagnosis plus a manifestation code may now have a single combination code -- but that combination code maps differently under V28's 115-HCC structure than the old pairing did under V24's 86-HCC structure. Without updated crosswalk logic applied at the point of coding, these shifts go unnoticed until a retrospective RADV audit surfaces them.

The Documentation Gap V28 Exposes

The core issue is not coder negligence -- it's documentation specificity that clinicians historically had no incentive to provide. A cardiologist documenting "heart failure" was adequately specific for clinical purposes, for V24 risk scoring, and for most payer contracts. Under V28, heart failure without systolic/diastolic distinction and without acuity qualifier (acute, chronic, acute-on-chronic) may map to a lower-value HCC or none at all. The documentation gap is structural, not incidental, and it exists across thousands of patient encounters every quarter.

The conditions where V28 specificity requirements bite hardest include:

• Chronic kidney disease (stages 1-5 vs. unspecified)

• Heart failure (systolic vs. diastolic; acute, chronic, acute-on-chronic)

• Diabetes mellitus with complications (type, specific complication, laterality)

• Dementia with or without behavioral disturbance (V28 added new HCC granularity here)

• Morbid obesity with specific comorbidities (BMI documentation now affects HCC mapping)

Where Agentic AI Fits in the V28 Workflow

Traditional computer-assisted coding tools are code-lookup engines. They identify a diagnosis string in a note and return the most probable ICD-10 code. They do not reason about HCC impact, they do not track whether a condition has been addressed with sufficient MEAT criteria (Monitoring, Evaluation, Assessment, Treatment) to survive an audit, and they do not flag when a documented code maps to no HCC under V28 but an adjacent, equally supportable code would.

Agentic AI systems operate differently. They ingest the full clinical note, the patient's longitudinal problem list, the current encounter context, and the V28 HCC-to-ICD crosswalk simultaneously. When they surface a code suggestion, they attach a chain of reasoning: the clinical evidence in the note that supports the specificity level, the HCC that code activates, the RAF impact, and a MEAT documentation assessment. The coder or CDI specialist sees not just "code X" but "code X because the note documents stage 3 CKD in the assessment, the creatinine trend in the labs supports it, and MEAT is satisfied by the medication review in the plan -- this maps to HCC 137 under V28 for a RAF of 0.289."

That reasoning layer is what makes agentic AI substantively different from prior-generation tools for the V28 environment. The coding decision is defensible in a RADV audit because the AI preserved the evidence chain, not just the code.

The April 2026 Code Additions: A Practical Opportunity

The 487 new ICD-10-CM codes effective April 1, 2026 are not uniformly a threat. Some of them represent a coding opportunity. CMS designed V28 to reward documentation specificity, and some of the new codes capture clinical granularity that previously had no dedicated home. New injury combination codes in Chapter 19, for example, consolidate fracture type, laterality, and episode of care into a single code -- and those combination codes map more cleanly to procedure codes and to post-acute HCCs than the multi-code sequences they replace. For MA plans and their coding partners, the April additions are a reason to run a gap analysis against the current active patient population, not just to update the chargemaster.

Raapidinc's analysis, published in April 2026 on their Medicare Advantage coding blog, documented specific HCC mapping shifts triggered by the April additions and estimated that plans without proactive coding review face millions in RADV penalty exposure from undetected specificity mismatches. That exposure is auditable, retrospective, and compounding.

Building an Agentic AI Workflow for V28 Compliance

The practical implementation of agentic AI for V28 risk adjustment involves three stages. The first is prospective: the AI reviews scheduled patient encounters before the visit, surfaces known chronic conditions that haven't been addressed in the current coding year, and flags where prior-year codes used V24 logic that needs V28 reclassification. The second is concurrent: during the encounter or immediately after, the AI reviews the generated note against V28 HCC criteria and recommends specific language additions to meet MEAT standards for each condition. The third is retrospective: a periodic sweep of closed encounters, identifying cases where a valid V28-specific code was left on the table and generating a query to the provider for clarification and amendment.

None of these stages requires replacing coders. They require coders who know how to use agentic AI as a reasoning partner -- reviewing its evidence citations, accepting or rejecting its suggestions with clinical judgment, and documenting the decision trail for audit purposes. The role shifts from code lookup to code validation, which is a higher-skill, more defensible, and ultimately more valuable function.

The 2026 Stakes

Medicare Advantage enrollment passed 35 million in 2025 and continues to grow. Every percentage point of average risk score reduction translates to billions in aggregate capitation payments. The 3.12% projected reduction from V28 is a floor, not a ceiling -- plans with documentation gaps will see larger reductions. Plans that address those gaps systematically, with agentic AI as part of the workflow, will not only protect current revenue but position themselves for the RADV audit environment that CMS is intensifying. The April 2026 code additions are a forcing function: the crosswalks changed, the HCC mappings shifted, and plans that don't update their coding logic will leave money on the table in a compliance framework that has no tolerance for ambiguity.

If your organization is navigating the V28 transition and needs a system that reasons about HCC impact -- not just generates codes -- explore Medikode's automated medical coding platform. It's built to handle the specificity requirements of V28 and the April 2026 code additions without adding manual review burden to your coding team.