AI is now reading medical records at a scale no coding team could match. That capability is genuinely useful for risk adjustment, but it raises a question every compliance officer and VP of Risk Adjustment should be asking: is the system finding conditions that were documented, or conditions that could be inferred? The difference determines whether a submitted diagnosis survives a RADV audit, and with CMS preparing to audit every eligible Medicare Advantage contract on samples of up to 200 records, that difference is now material.

Two Very Different Things

A documented HCC is a condition the treating clinician recorded during a qualifying face-to-face encounter, with evidence that it was monitored, evaluated, assessed or treated. It is in the note, in the clinician's words, and it can be coded directly.

An inferred HCC is a condition that the evidence suggests may exist, but that the clinician did not document as a diagnosis in that encounter. Common examples:

  • A lab value in a diagnostic range (eGFR consistent with CKD stage 3, for instance) without a corresponding diagnosis.
  • A medication that implies a condition, such as insulin with no diabetes documented in the visit, or an anticoagulant with no documented indication.
  • A condition documented in a prior year, a specialist letter or a hospital discharge summary but not addressed in the current encounter.
  • A BMI that meets a clinical threshold without documentation of the associated diagnosis.

Inferred evidence is valuable. It points to care gaps, underdiagnosed conditions and patients who need attention. But it is not a codable diagnosis. Coding guidelines require that diagnoses be based on the provider's documentation, and coders cannot assign a diagnosis from lab values, medications or other clinical indicators alone. RADV reviewers look for a diagnosis documented by an acceptable provider in an acceptable record, not for evidence that a diagnosis was plausible.

Where AI Chart Review Goes Wrong

The risk in AI chart review is not that the technology is inaccurate. It is that it can be pointed at the wrong objective. A system optimized to maximize HCC capture will naturally surface inferred conditions alongside documented ones, and if the workflow does not clearly distinguish them, inferred conditions can end up on claims or in chart review submissions.

The regulatory environment makes this especially dangerous. The OIG has reported that diagnoses appearing only on health risk assessments and chart reviews, with no follow-up care, accounted for an estimated $7.5 billion in 2023 MA payments, and it has repeatedly identified chart review as a source of diagnoses that lack support. A chart review program that adds diagnoses without adding the corresponding clinical documentation looks, from an auditor's perspective, exactly like the pattern OIG has warned about.

There is also an extrapolation problem. Under the RADV rule CMS finalized in 2023, findings from audited samples can be extrapolated across a contract beginning with payment year 2018. If AI-assisted review systematically adds a category of weakly supported diagnoses, the error does not stay small. It gets multiplied.

How Responsible AI Chart Review Should Work

The right design principle is simple: AI surfaces evidence, people confirm diagnoses. In practice, that means four things.

1. Classify every finding

Each AI finding should be labeled as documented (the condition is stated and supported in the encounter note) or inferred (the evidence suggests a condition that is not documented as a diagnosis). The label should be visible to every person who touches the record, and the two categories should follow different workflows.

2. Route documented findings to coders

Documented findings go to certified coders for validation. The coder confirms the diagnosis, checks specificity and verifies that the note shows active management. The AI's job is to make the coder faster and more consistent, highlighting the supporting language and flagging gaps, not to replace the coder's judgment.

3. Route inferred findings to clinicians

Inferred findings become clinical questions, not codes. They should go to the treating provider as a compliant query, prospectively in a pre-visit summary for the next encounter, or concurrently while the note is still open. The clinician evaluates the patient and decides whether the condition exists. If it does, the clinician documents it, with evidence of management, in a qualifying encounter. If an addendum is appropriate, it is signed by the clinician and reflects care actually provided.

4. Keep the evidence trail

Every decision should be traceable: what the AI found, what evidence it cited, who reviewed it, what they decided and why. That trail is what lets an organization explain its process to an auditor and demonstrate that diagnoses were confirmed by qualified people rather than generated by software.

Deletes Matter as Much as Adds

A responsible program uses AI to find unsupported diagnoses as well as missed ones. If the AI reviews a note and finds a coded condition with no evidence of management, that is an overpayment risk. Programs that only use technology to add diagnoses invite the conclusion that the program exists to raise risk scores. Programs that also identify and remove unsupported codes demonstrate good faith and reduce RADV exposure directly.

Questions to Ask About Any AI Chart Review Program

  • Does the system distinguish documented from inferred conditions, and is that distinction visible in the workflow?
  • Can an inferred condition ever reach a claim or chart review submission without clinician documentation?
  • Does the system identify unsupported codes for deletion, not just missed codes for addition?
  • Does every finding cite the specific source text or data it relies on?
  • Who has final authority on each diagnosis, and is that recorded?
  • How are support rates measured, and how do they compare with a RADV-style sample?

Defensibility Is the Point

Curitics' AI chart classification is built around this separation: it labels each HCC as documented or inferred, sends documented findings to coders for validation and turns inferred findings into provider queries that end, where appropriate, in signed addenda. Prospective, concurrent and retrospective review share the same evidence trail, with patient-level V28 scoring showing what each confirmed decision is worth.

The Takeaway

AI will make chart review faster and more thorough. Whether it makes risk adjustment more accurate depends on how it is governed. Organizations that keep the line bright between what the record says and what the data suggests will capture legitimate revenue and defend it. Those that blur it are building audit findings into their own submissions.