6.09%
of Medicare Advantage payments, about $23.7 billion, were estimated improper in FY2025. Overpayments, mostly diagnoses the record did not support, were 5.51%.
CMS · FY2025 (CY2023 payments) · gross rate · "improper" is not fraud
Every code has to hold up.
Coders and auditors have to reach a defensible HCC decision when the documentation in front of them is incomplete, ambiguous or conflicting, and get it right in both directions.
Two ways to get it wrong. They are not the same size.
A condition that is documented and managed but never coded understates how sick the patient is. That is an accuracy failure, not a revenue opportunity. In one plan's audited sample of 200 patients, 108 of the 688 HCCs it coded were not supported by the record. The same records supported 11 more that were never coded, found only while checking the coded ones. CMS's national measure does not count conditions a plan never submitted.
HHS-OIG · one Puerto Rico plan's random sample of 200 enrollees · payment year 2017
Both directions are the standard. The evidence weighs far more heavily on one side.
Finding without defending is what regulators police.
- 2016 encounter data · HHS-OIG
99.3% of chart reviews added diagnoses; 0.7% deleted them.
OIG did not determine whether the added diagnoses were supported.
- February 3, 2026 · HHS-OIG compliance guidance
Lists, among conduct federal investigations have revealed, failing to remove codes that chart reviews showed were unsupported, and prompting physicians, including through AI-generated prompts, to add diagnoses.
Voluntary guidance.
- 2026 · U.S. Department of Justice
Kaiser Permanente affiliates ($556M, January 14), Aetna ($117.7M, March 10–11) and Independence Blue Cross ($22.5M, September 30) settled False Claims Act allegations over Medicare Advantage risk adjustment.
Allegations only; no determination of liability.
The standard is written down.
CMS
Every diagnosis submitted for risk adjustment must be documented in the medical record as the result of a face-to-face visit and coded to the ICD-10-CM guidelines. When a plan's own review finds codes that fail, it must delete them.
ICD-10-CM
No outpatient diagnosis documented as "probable", "suspected" or "rule out"; a coexisting condition only when it affects care; the code follows the provider's own statement, with a query when the record conflicts.
RADV
Certified medical-record coders read the record as written, with a third review when the first two disagree. CMS has said that decisions that could lead to an overpayment finding will be made by human certified coders. A January 2026 commitment about a planned tool.
What changed.
- CMS-HCC V28
115 payment HCCs, up from 86. On the FY2022/23 code set, the ICD-10-CM codes that map to a payment HCC fell from 9,797 to 7,770.
They remain valid codes.
- Payment year 2024
CMS's audit lists cover every eligible Medicare Advantage contract: 590, against 58 for 2018. Those audits began on August 28, 2026.
Selection is routine, not a finding.
- 2026 dates of service → 2027 payment
Diagnoses from chart reviews not tied to a recorded encounter are excluded; estimated −1.53% average effect on risk scores.
Delayed to the 2027 midyear run; linked chart reviews still count.
In Puerto Rico, more rides on the risk score.
95%
of Puerto Rico's Medicare beneficiaries with Part A and Part B were in Medicare Advantage in March 2026, against 55% nationally.
Every Puerto Rico Medicare Advantage contract is on CMS's audit lists for payment years 2019, 2020, 2021 and 2024, and HHS-OIG found codes the record did not support at each of the three Puerto Rico organizations it audited. So the Puerto Rico case is defensibility, not recapture.
The headline and the last sentence are our reading of these figures, not a finding of any source.
What AI research hasn't tested yet.
Published evaluations of AI medical coding mostly measure agreement with codes humans assigned. A few 2026 studies test whether a model can judge that an inpatient code is supported by the note. In a search on October 12, 2026, we found none that evaluates whether AI judges an HCC to be supported under CMS's RADV standard.
A dated statement about our search, not a claim that AI cannot do it.
Be first in line.
Neurona 28 is in development. Join the waitlist and we'll write when it's ready for you.
Join the waitlist