Finding Codes Was Never the Hard Part
NLP has been able to identify diagnosis mentions in clinical notes for years. The technology to scan a chart and surface potential HCCs isn’t new, and it isn’t rare. Dozens of systems on the market can open a medical record and flag coded conditions. That’s the finding part, and it’s largely a solved problem.
The proving part is where programs fail. Finding a diagnosis in a chart and proving that the diagnosis meets CMS’s evidentiary standard are two entirely different operations. The first requires pattern recognition. The second requires clinical evidence validation against MEAT criteria (Monitoring, Evaluation, Assessment, Treatment), documentation quality assessment, and an explainable decision trail that an auditor can follow.
The OIG’s BCBS Alabama audit (A-07-22-01207, March 2026) illustrated this gap. 91% of sampled enrollee-years had unsupported diagnosis codes. The codes were found in the charts. They just couldn’t be proven. History-of conditions were coded as active. Chronic diseases appeared in problem lists without evidence of current management. The technology found the diagnoses. Nobody validated whether the documentation could defend them.
The Validation Gap in Current Technology
Most tools in the market were designed around the finding function. They ingest clinical notes, apply NLP or machine learning to identify diagnosis mentions, and present a list of potential HCCs to the coder. Some flag confidence levels. Some highlight relevant text passages. But few systematically evaluate whether the documentation meets the MEAT standard that CMS applies during audits.
That gap creates a dangerous workflow. The coder sees a list of AI-recommended codes, reviews them under time pressure, and submits the ones that seem reasonable. There’s no structured MEAT validation step. There’s no documentation quality score. There’s no audit trail showing why each code was deemed defensible. The submission is based on the coder’s judgment working from an AI recommendation, and when the audit arrives, neither the coder nor the AI has produced the evidence CMS is looking for.
The Aetna DOJ settlement ($117.7 million, March 2026) and Kaiser settlement ($556 million) both involved programs where coding volume outpaced validation rigor. The tools processed charts efficiently. The validation architecture was inadequate.
Novafork: The Ultimate Guide to This Innovative Tool
What Proof-Ready Tools Do Differently
Tools built for the current environment treat validation as the primary function, not a secondary check. When the system identifies a potential HCC, it simultaneously evaluates the documentation against MEAT criteria. It maps specific sentences in the note to specific evidentiary elements. It flags where evidence is strong, where it’s ambiguous, and where it’s absent. The coder receives a validated recommendation rather than an unvalidated suggestion.
Two-way capability completes the picture. The system evaluates not just potential adds but also existing submissions. It identifies codes that were previously submitted but now lack adequate documentation support, whether because the condition resolved, the provider stopped documenting active management, or the original chart review missed a MEAT gap. Deletions get the same evidence trail as additions.
The output isn’t a list of codes. It’s a documented, evidence-linked coding package where every recommendation traces back to clinical language in the note and every gap is identified before submission.
Closing the Gap
The distinction between finding codes and proving codes should drive every technology evaluation in risk adjustment today. Plans that invested in tools optimized for finding are now discovering that those tools don’t produce the evidence required to defend what they found. The enforcement environment has made this gap financially consequential.
Any Risk Adjustment Tool deployed in 2026 needs to close this gap by design, not through workarounds or manual compliance layers bolted on after the fact. MEAT validation, two-way review, and explainable evidence trails need to be core capabilities, not optional modules. The plans that select tools built for proving will outperform those still relying on tools built for finding alone.
