OIG's first MA compliance guidance since 1999 names AI EHR prompts: verify every AI-surfaced diagnosis

A diagnosis can reach your queue two new ways that didn't exist a few years ago. An ambient AI scribe drafts it into the note from the visit audio, or an EHR prompt suggests it at the point of care. Neither one is documentation. On February 3, 2026, the HHS Office of Inspector General made that distinction a compliance problem.

OIG published its Medicare Advantage Industry Segment-Specific Compliance Program Guidance that day, the first compliance guidance the office has written specifically for MA since 1999. Its risk-adjustment section reads like a list of the shortcuts coders already push back on, and now it has AI in it by name.

What OIG actually named

The risk-adjustment section (part C, around page 20) lists conduct OIG treats as fraud and abuse. Four items land on the coder's desk, in OIG's own words:

• "querying physicians via electronic medical record platforms (including prompts generated by artificial intelligence algorithms) ... to add risk-adjusting diagnoses that patients did not have"
• "using chart reviews to identify additional diagnoses that increased risk scores inappropriately"
• "failing to remove diagnosis codes previously submitted to CMS when chart reviews provide information that those codes were unsupported or otherwise invalid"
• "providers submitting diagnoses that were not supported by the enrollees' medical records"

Read together, they point one way. OIG treats add-only chart review as an exposure, and it treats removing an unsupported code as a requirement, not a courtesy. OIG separately flags AI in utilization management, where an algorithm "may not be compliant ... if it determines coverage based on a larger data set instead of the individual patient's medical history." That is the prior-auth angle, adjacent to your desk, not the coding one.

The error you won't catch is the omission

The EHR prompt suggests a diagnosis. The ambient scribe writes the note the diagnosis has to be supported in. The failure mode that should change how you read a scribe note is the one you can't see: omission. In a simulated-encounter study across ambient platforms, omissions were the most common error, about 76% of all errors found. A condition gets addressed in the room, the scribe leaves it off the page, and nothing on the page looks wrong. The same research documents meaning-reversal errors, where the note flips what happened: one platform logged a patient as "alert" in a case involving altered mental status, another dropped a medication allergy the patient had reported. Reported overall error rates for modern ambient scribes run around 1% to 3%, small in the abstract and consequential in a chart.

Keep that error research separate from the efficiency research, because they answer different questions. A randomized trial in NEJM AI (Lukac et al., 238 outpatient physicians across 14 specialties) measured documentation time and burnout, not accuracy. It found Nabla cut time-in-note about 9.5% while Microsoft DAX Copilot showed no significant change. The time savings are small and product-specific. That trial says nothing about error rates, and you shouldn't cite it for any.

Two AI inputs, one check

Both inputs do the same thing to your chart: they add. Here is each one mapped to the check it forces.

First, does it even map to an HCC?

Before you spend a minute chasing documentation, confirm the AI surfaced something that risk-adjusts at all. A prompt or a scribe will float diagnoses that carry no HCC in the payment-year model, and those never move a risk score no matter how cleanly they are documented. Run the code through the ICD-10-to-HCC crosswalk and check it against the V28 model before you treat it as a capture worth defending. If it doesn't map, it isn't a risk-adjustment question, and your verification time belongs on the codes that are.

The guidance cuts against "more codes"

Every one of those inputs pushes more diagnoses onto the chart. OIG's rule cuts the other way. An AI-prompted HCC with no record support is precisely what the guidance says has to come off, and failing to remove it is the violation OIG describes. OIG doesn't ban the tools. The unsupported code is the problem, whether a coder, a chart review, or an algorithm surfaced it.

It's the same check you already run. The only new thing is where the diagnosis came from, so every AI-surfaced diagnosis still needs its own MEAT before you submit the HCC. Walk it through the MEAT criteria and confirm the supporting evidence is in the record at the date of service. If it isn't, the code doesn't move the RAF, because it shouldn't be on the claim at all.