The audit will tell you which AI was ready

Risk adjustment payments fund the care Medicare Advantage plans provide to their sickest members, and CMS has tightened how it verifies that funding is earned. Health plans that have internalized the shift are already making different decisions about who they trust to power their risk adjustment programs.

RADV now runs quarterly, and the settlements are getting larger

RADV audits are now occurring quarterly and are expected to continue through 2027 as CMS works through its backlog. The PY 2020 methodology was designed to calculate error rates from unsupported HCCs found in audit samples and extrapolate those findings to the applicable sampling frame, significantly increasing recoupment exposure. A federal court vacated the 2023 RADV final rule in 2025, and CMS has appealed the decision.

Federal scrutiny continues to intensify. Kaiser Permanente affiliates settled False Claims Act allegations for $556 million in January 2026, the largest Medicare Advantage fraud settlement of its kind. For health plan executives managing Medicare Advantage books, coding accuracy coding accuracy now belongs on the executive agenda as a financial and operational priority.

Explainability is now the regulatory floor

The market conversation about AI in risk adjustment has narrowed to a single word, explainability. An algorithm that keeps its clinical reasoning hidden generates audit risk, and health plans operating under intense RADV scrutiny pay a steep price for opacity in their coding programs. My concern is how quickly the premise has collapsed into marketing language, and how many procurement decisions rest on a criterion that stops one question too early.

Linking every HCC to encounter-based clinical evidence with full MEAT documentation is the regulatory floor in 2026, and any platform competing seriously in Medicare Advantage must clear it. In a RADV audit, a DOJ inquiry, or a CMS data validation review, the outcome turns on whether that explainability withstands the scale, speed, and data complexity of a real enterprise risk adjustment program. The test is materially different, and

The real test is whether explainability withstands enterprise scale

A Medicare Advantage risk adjustment program at scale is managing millions of member encounters across fragmented provider networks, with clinical documentation arriving through EMR integrations, direct retrieval, fax, and field-based medical record access, across retrospective, prospective, and concurrent workflows running simultaneously. The regulatory scrutiny spans every one of those data sources, under audit timelines that have compressed considerably in the last eighteen months. A platform that performs cleanly in a controlled demonstration environment and a platform that performs reliably under those conditions are two distinct things, and health plans evaluating AI in 2026 should be demanding evidence of the latter, grounded in production performance from live customer deployments.

At Reveleer, we have lived this test across the patient lives entrusted to our platform. Reveleer processes 200+ million pages of chart records annually across 387,000+ providers and 75+ HIEs, operating across more than 10 million covered lives. The customers who have run RADV audits on our platform, managed DOJ-adjacent scrutiny, and navigated the PY2020 extrapolation cycle have remained with us at a 97% retention rate, a number earned through the hardest compliance environment Medicare Advantage has produced in a decade

Health plans past the evaluation stage already know which technology partners proved defensible under regulatory scrutiny."
_{- Jay Ackerman, President & CEO, Reveleer}

‍

Audit-defensible AI requires transparency and scale at once

Audit-defensible AI requires two things to be true simultaneously. The reasoning must be transparent at the level of individual coding decisions, and the platform must be capable of delivering that transparency at enterprise volume, under real audit pressure, without degrading the accuracy or the evidence quality that makes the output defensible in the first place. Health plans that are past the evaluation stage and into execution already know which of their vendors proved defensible under regulatory scrutiny. For those still making that determination, the audit will answer the question one way or another, and 2026 punishes delay. This same standard applies to Reveleer.

We built EVE GenECS for audit-defensible AI at production scale

Reveleer measures itself against the same test this article describes, and the answer starts with EVE GenECS. GenECS is the generative AI engine inside Reveleer NextGen Risk Adjustment, built for retrospective coding and part of the EVE™ portfolio of AI capabilities. It runs on infrastructure Reveleer built from the ground up and controls end to end with:

• A knowledge base trained on validated coded chart history across Medicare Advantage and ACA
• A retrieval pipeline engineered for risk adjustment coding
• A Reveleer-controlled scoring engine that ties every recommendation to a specific clinical excerpt, the page it came from, and a confidence score

The evidence trail makes GenECS auditable at the source, in a market still leaning on black box AI, and getting that evidence right protects the funding health plans need to care for their sickest, most complex members. Competitors without comparable data volume fall short of that knowledge base, and it compounds with every chart Reveleer's coders validate. Reveleer governs the pipeline from data through release, so coding output changes only on Reveleer's timeline, with advance notice, and Reveleer stays accountable for the coding record when a chart reaches a RADV audit.

The coder workflow stays the same in our expert risk adjustment coding solution. Coders open a Chart, review a tighter recommendation set, and accept or reject each code with the supporting evidence already attached, all in the interface they already use. GenECS is one factor in a coding program, alongside staffing, coding guidelines, coder proficiency, chart quality, and data completeness. Bringing generative AI into a clinical coding process the right way means proving it at real volume first, then migrating each customer on a managed, per-project basis as the evidence supports it.

If your current risk adjustment platform remains untested at enterprise scale under live RADV pressure, that is a conversation worth having before CMS initiates it for you.

‍

‍

RADV audit FAQs

How has RADV audit exposure changed for Medicare Advantage plans in 2026?

RADV audits run on a quarterly cadence through 2027 as CMS works through its backlog, then shift to roughly one audit cycle per payment year, extending CMS's review reach to all eligible Medicare advantage contracts and enrollee samples ranging from 35 to 200 based on contract size. The PY 2020 methodology was designed to use errors found in a sample to estimate overpayments across the applicable sampling frame, allowing a small number of unsupported diagnoses to create six-figure recoupment exposure rather than a chart-by-chart correction. A federal court vacated the 2023 RADV final rule in 2025, and CMS has appealed the decision governing enforcement of those extrapolated recoveries.

What financial exposure does extrapolation create for a health plan's leadership?

Extrapolation converts a sample-level finding into a cohort-level liability. Leadership must reserve against a repayment figure far larger than the sampled charts alone would indicate. The extrapolated liability now sits on the CFO's balance-sheet planning agenda alongside the compliance team's audit calendar.

What separates explainable AI coding from a black box?

Explainable AI coding links every suggested diagnosis code to the specific encounter and chart evidence that supports it, letting a coder or auditor trace the reasoning behind each recommendation. A black box model returns a code and a confidence score disconnected from the clinical documentation, leaving the coder without evidence to defend when CMS asks. GenECS follows the first standard. Every recommendation includes the clinical evidence behind it, credible by design instead of by confidence score alone.

What is EVE GenECS?

GenECS is the generative AI engine inside Reveleer Risk Adjustment, built for retrospective coding and part of the EVE™ portfolio of AI capabilities. It ties every recommendation to a specific clinical excerpt, the page it came from, and a confidence score, and it runs alongside ECS, the version-locked engine, for customers whose agreements restrict generative AI.