Medicare Advantage risk-adjustment coding teams have spent the better part of a year working without a settled rulebook. A federal court vacated CMS’s RADV extrapolation methodology in September 2025, CMS appealed that decision to the Fifth Circuit, and as of this spring the case was still being briefed with no ruling in sight. For coders and compliance staff, that means the audit standard their diagnosis codes will eventually be measured against is not fully known yet. It also means the coding platforms and workflows built to support them can’t just optimize for throughput — they have to be able to withstand scrutiny under whatever standard lands.
The RADV Uncertainty Coding Teams Are Living With
Risk Adjustment Data Validation (RADV) audits are how CMS checks whether the diagnosis codes a Medicare Advantage plan submitted are actually supported by the medical record. When they’re not, CMS can claw back payment — and historically it has used extrapolation, applying an error rate found in a small audited sample to a plan’s entire membership, to calculate what it recovers. That extrapolation methodology is exactly what’s been in legal limbo.
What the Humana Ruling Actually Changed
In Humana Inc. v. Becerra, the U.S. District Court for the Northern District of Texas ruled on September 25, 2025 that CMS had changed its core justification for eliminating a long-standing adjustment factor in its 2023 RADV final rule without following proper notice-and-comment procedure, and vacated the rule. CMS filed its notice of appeal on November 21, 2025, and by late March 2026 had filed its opening brief with the U.S. Court of Appeals for the Fifth Circuit in Humana Inc. v. Becerra, No. 25-11293 — according to Crowell & Moring’s client alert on the appeal. The practical effect: health plans and the coding teams that support them still don’t know exactly how CMS will calculate extrapolated overpayments once the litigation resolves.
That uncertainty is compounding, not resolving. When CMS finalized its CY2027 Medicare Advantage rate notice on April 6, 2026, it opted to delay a planned update to the underlying risk-adjustment model and continue using the existing 2024 model — built on 2018 diagnosis data and 2019 expenditure data — for another payment year. Coding teams are effectively being asked to stay accurate against a moving, and in some respects frozen, target at the same time.
Why Coding Speed Alone No Longer Cuts It
Most of the automated and AI-assisted coding tools on the market today are still primarily judged on throughput and raw accuracy against a coder’s gold-standard label. Those are the wrong metrics for a risk-adjustment environment defined by RADV appeals, extrapolation uncertainty, and continued OIG audit activity. A code that’s technically “correct” in the sense that it matches what a human coder would have assigned isn’t worth much to a compliance team if nobody can reconstruct, six months or two years later, exactly which chart documentation supported it and why.
Agentic AI changes what’s possible here, but only if it’s built with that reconstruction requirement in mind from the start. An agentic system doesn’t just classify a chart and output a code — it can work through a documented reasoning chain, flag ambiguous or thin documentation for human review instead of guessing, and preserve the evidence trail that produced its output. That trail is the difference between a coding platform that helps in an audit and one that becomes a liability in one.
What Audit-Ready Agentic AI Coding Looks Like
For Medicare Advantage risk-adjustment coding specifically, “audit-ready” isn’t a marketing phrase — it maps to a specific set of technical capabilities a coding system either has or doesn’t:
- Chart-to-code traceability: every submitted HCC code links back to the exact clinical documentation excerpt that supports it, not just a confidence score.
- Confidence-scored routing: low-confidence or thinly documented diagnoses route to a human coder automatically instead of being auto-submitted.
- Version-controlled logic: when the coding model or ruleset changes, the platform retains a record of which version produced which code, so a two-year-old submission can still be explained.
- Deletion and correction history: if a code is later withdrawn or corrected, that action — and the reason for it — is logged, addressing the exact kind of chart-review-but-don’t-correct pattern that has drawn regulatory scrutiny elsewhere in risk adjustment.
- Model-version alignment with CMS’s current HCC model: the platform’s logic reflects whichever model year CMS is actually using for payment, not last year’s assumptions.
Why This Matters More With the Rules in Flux
When the extrapolation standard itself is under appeal, plans can’t simply wait for certainty before acting — RADV audit cycles continue regardless of where the litigation stands. Groom Law Group’s analysis of the underlying ruling notes that it remains unclear how CMS will approach RADV audits going forward, which leaves plans managing audit risk under multiple possible future rule sets at once. A coding platform that can document its reasoning under any of those scenarios is a materially safer bet than one that can only show its work under today’s assumptions.
How Coding Teams Should Prepare Now
Waiting for the Fifth Circuit to rule isn’t a coding strategy. In the meantime, risk-adjustment coding teams and the compliance staff who oversee them should be pressure-testing their current tooling against a few concrete questions: can we produce, for any submitted HCC code from the last three years, the specific documentation that supported it, within minutes rather than days? Do we have a defensible answer for why a code was or wasn’t corrected after a chart review flagged it? And is our coding logic tied to the risk-adjustment model CMS is actually using for the current payment year, not an assumption baked in when the tool was configured?
Vendors selling coding automation should be able to answer those questions concretely, not with a general accuracy percentage. If they can’t, that’s a gap worth closing before an auditor — not a sales conversation — is the one asking.
Getting Coding Right Before the Audit, Not After
The RADV extrapolation fight will eventually resolve, but the underlying expectation isn’t going away: risk-adjustment coding has to hold up under scrutiny, not just move fast. That’s the standard Medikode’s automated medical coding platform is built around — coding that’s fast because it’s accurate, and defensible because every code carries its own documented reasoning back to the chart.