Can AI rescue physicians from Medicare's documentation arms race?

Medicare risk adjustment was designed as a technical correction to the original Medicare Advantage (MA) payment formula, so that capitated payments to MA plans would track the expected cost of their enrollees. Medicare Advantage plans that disproportionately enroll sicker patients with higher predicted spending would receive higher payments, while plans whose membership is healthier and lower‑cost would receive reduced payments.

In theory, that is straightforward. In practice, over the past 15 years, I’ve watched it drift from a clinically grounded concept into a national game of “how many boxes can you check,” with billions of dollars riding not on how sick patients actually are, but on how aggressively plans and vendors can mine charts for billable diagnoses.

I’ve lectured to almost 20,000 clinicians and medical staff on risk adjustment over the past 15 years. I have seen this subject matter deteriorate with little regard for patient prognosis. Instead, it’s a steady drumbeat of reminders to “refresh hierarchical condition categories [HCCs] annually,” prompts to re‑label “history of” problems as active conditions and offers from third‑party vendors to comb through notes and problem lists for any International Statistical Classification of Diseases, Tenth Revision (ICD‑10) code that might map to an HCC. A payment tool originally designed to neutralize selection bias has evolved into what can only be called a documentation arms race.

V28: Fixing diabetes gaming, creating new blind spots

The full phase‑in of the Centers for Medicare & Medicaid Services (CMS) HCC Version 28 (V28) model for Medicare Advantage (MA) in 2026 tries to address some of the worst distortions, with diabetes being a prime example. Under V28, diabetes codes with or without complications now carry the same risk adjustment factor (RAF) value, reducing the incentive to “upgrade” every diabetic to have a complication simply to boost scores. That is a real improvement in terms of curbing some of the more blatant gaming. But it also exposes a deeper problem: Not all patients with diabetes are the same, clinically or in terms of risk, and a flat RAF treatment for diabetes can obscure meaningful differences in severity and prognosis.

When box‑checking replaces patient care

When risk adjustment becomes box‑checking, it stops serving patients and starts serving paperwork. If we want risk adjustment to reflect real risk instead of coding games, we need to be honest about how far we’ve drifted — and then acknowledge that, despite its imperfections, modern artificial intelligence (AI) may give us a path to do this very differently. Properly designed, AI could help us move away from manual, code‑by‑code score chasing and toward models that capture true patient complexity from the full clinical record, aligning payment more closely with reality, instead of with whoever can click the most boxes.

How we got a box‑checking RAF system

On paper, RAF scores are simple. A higher RAF score means a higher expected cost, so the plan gets a higher capitated payment. The catch is that only certain diagnoses “count,” and only if they are documented in coded claims in the past year.

The incentive is “more codes,” not “more accuracy.” If a condition is documented in the note but not coded on a claim, it might as well not exist for payment purposes. The primary output isn’t better care: It’s more codes.

Two patients with similar diseases and utilization patterns can have very different RAF scores simply because one clinician writes sparse problem lists while another uses templates that pull in every historical diagnosis. The current system pays more attention to who is best at playing the documentation game than to which patients are actually at higher risk. It shifts documentation toward higher-paying diagnoses rather than more accurate ones, undermining the integrity of risk adjustment and blurring the line between legitimate coding and revenue optimization.

Clinicians experience risk adjustment as burden and distortion

Risk adjustment doesn’t typically show up as a tool that helps a primary care physician manage a complex panel. It shows up as queries and nudges to add codes that have little to do with that day’s clinical decision-making. Every minute spent on “RAF documentation optimization” is a minute not spent on medication reconciliation, care coordination or genuine shared decision-making.

The result is a paradox. We have a highly engineered risk‑adjustment apparatus that is extremely sensitive to whether a specific ICD‑10 code was clicked in a specific calendar year, and surprisingly insensitive to what clinicians actually know about the patient’s prognosis, functional status or social risk.

The hidden costs of a documentation arms race

The current RAF system doesn’t just misallocate dollars. It also shapes the industry’s behavior in ways that are bad for patients and primary care. Capital flows to coding, not care. When the highest return on investment in a business line is more accurate chart-mining, not better prevention or chronic care management, it is rational for plans to spend heavily on coding infrastructure and vendors. That money comes from somewhere, and too often it is diverted from frontline clinical care or left as profit, rather than reinvested.

What AI can see that boxes cannot

The current RAF/HCC approach uses only a fraction of information — demographics plus a limited list of diagnosis categories derived from claims. Modern AI can do this in a more direct and clinically meaningful way. AI could estimate risk instead of asking humans to reverse‑engineer risk through boxes by using the full record and not just the bill. Today’s RAF is mostly driven by claims. AI models can ingest longitudinal electronic health record data, trends in laboratory results, imaging, medications, utilization patterns, social risk flags and the narrative in notes. They can see that a patient with “mild” chronic obstructive pulmonary disease who has three emergency room visits and two steroid bursts is at greater risk than another patient with an identical diagnostic label but stable spirometry and no exacerbations. The current system treats them the same if the codes match.

An AI‑assisted system could assign each patient a risk index that updates as their health evolves. Risk would move when things actually change — new diagnoses, hospitalization, treatment intensification, functional decline — not just when a coder rediscovers last year’s HCC.

If risk scores are derived from a broad, validated model that looks at overall utilization patterns and clinical evidence, there is less payoff in simply adding one more HCC code. Yes, any system can be gamed, but a model that sees 50 to 100 data elements per patient is harder to manipulate than one that revolves around a few dozen code categories.

Designing AI‑enabled RAF that clinicians and regulators can trust

For an AI-assisted risk-adjustment system to be credible, it needs explicit design principles and guardrails with transparency over mystery. Regulators could specify the inputs and broad logic of any approved model. Health plans or CMS could publish model documentation, much like drug labels, so everyone knows how the risk is being calculated.

Clinicians will never trust a risk score that cannot be explained at the bedside. They may have more trust in an AI-driven RAF system that shows the “top drivers” of a given patient’s risk — recent admissions, abnormalities in laboratory results, polypharmacy, functional scores — rather than a cryptic number.

Any system that pulls from rich clinical data must be built on robust privacy safeguards and clear rules about who can access which elements and for what purpose. This is true today for coding vendors. It must remain true in an AI future.

Just as CMS audits RAF coding now, AI‑enabled risk models would need monitoring for drift, bias and gaming. That means regular validation against outcomes, comparison across plans and the ability to identify outliers whose risk scores seem out of sync with actual utilization.

If we get this right, the day‑to‑day experience of clinicians changes significantly. Instead of being urged to “add more diagnoses for accuracy,” physicians could see a dynamic risk indicator that helps identify which patients are most likely to decompensate, need outreach or benefit from care management. Payment and care management would finally be using the same language of risk, grounded not in box‑checking but in what the data actually show about the patient’s medical illness trajectory.

Moving on from a coding contest

We should be honest about where we are. Today’s HCC/RAF regimen is a sophisticated coding contest dressed up as risk adjustment. It measures who is best at checking boxes inside an arbitrary time window more than it measures who is actually sick. That is not a criticism of clinicians or plans. It is a predictable outcome of the rules that have been written.

None of this requires surrendering control to a black box. The real question is not whether to choose AI or humans, but how to define the roles each should play. If a medical condition is actively treated or significant enough to be addressed during a visit, it should remain the physician’s or other clinician’s responsibility to document and code it accurately. Providers and Medicare Advantage plans should be required to attest that artificial intelligence is not being used solely to generate documentation for the purpose of inflating diagnosis codes. The widespread adoption of AI in clinical workflows is inevitable, including its use in drafting notes and suggesting diagnoses, which creates a clear risk of further manipulation of RAF-based risk adjustment.

AI-driven risk scoring can still incorporate EMR notes and diagnosis codes, but will be relying on longitudinal, objective data — such as specialist utilization, emergency department visits, hospitalizations, imaging, laboratory results, and pharmacy refill patterns — rather than documentation alone. This approach is far less susceptible to manipulation and more accurately reflects true illness burden. By evaluating multiple parameters, AI can better assess disease severity, recognizing that the presence of a diagnosis alone does not capture clinical complexity or cost. In this role, AI informs decision-making without replacing clinical judgment or accountability.

The choice is not whether to use risk adjustment, but how to do it responsibly. We can continue a documentation-driven cycle that rewards coding intensity, or we can use modern tools to align payment with actual patient need. Risk reflects the trajectory of illness and the realities of care—not the number of HCCs captured by year’s end.

Robert Resnik, M.D., MBA, is a board-certified internal medicine physician practicing in Cary, North Carolina. He earned his medical degree from Eastern Virginia Medical School and completed his residency at East Carolina University. He received an MBA from Duke University.