Your coding tool chose the diagnosis. Your signature certified it.

A Blue Cross Blue Shield Association analysis estimates that hospital AI coding tools added $942 million to its plans' costs over two years, most of it from secondary diagnoses. Its example: more anemia diagnoses without a matching rise in transfusions. The hospitals say the tools simply capture patients' conditions more accurately. Neither side will concede, and the fight won't be settled by whose algorithm is smarter. It will be fought one claim at a time, in the medical record.

The clinician carries a risk the institution doesn't

Payer audits will work in two steps. First, the AI looks for patterns across thousands of claims, like Blue Cross's anemia finding: a diagnosis rising with no matching rise in the treatment that usually comes with it. Then the flagged claims get pulled for review, and the question becomes whether each patient's record shows the diagnosis mattered. Was the condition evaluated, treated, tested or monitored? A diagnosis listed with no assessment and no plan won't survive that review.

For the payer and the hospital, win or lose, this is a fight about money — a claim gets paid, denied or clawed back, and both sides move on to the next one. The clinician is in a different position. Their signature certifies that the record is accurate, and on a Medicare claim that certification runs, in many instances, to the federal government: overpayment findings tied to their name, False Claims Act scrutiny, even exclusion from federal programs. And only a clinician can put the proof in the record, because the coding tool can spot a possible diagnosis but cannot examine the patient. A clinician who accepts a diagnosis they never evaluated hasn't supported the claim; they've certified it.

Most of the time this won't be a deliberate choice. Organizations are adding documentation assistants to save clinicians time, and some of those tools work alongside the coding software, gently prompting clinicians to note the conditions it has flagged. No one sets out to upcode. The record simply drifts, one helpful suggestion at a time, and the hospital chose the tools, but the clinician may be the one answering for them.

What a defensible program looks like

A defensible program protects the people whose names are on the record. It starts with a simple rule: no clinician accepts what they didn't evaluate. AI prompts should read as questions, and accepting one should require the clinician's own assessment and plan.

Clinicians should also know what their tools do, including whether the documentation assistant is connected to the coding tool. Show them what's billed under their names, before and after the tools went live. If it changed and their patients didn't, find out why before a payer does.

Finally, give them a safe way to raise concerns with compliance, in writing, without consequence. "The algorithm suggested it" is an explanation, not a defense, and an organization that listened to its clinicians will be far better placed than one that didn't.

The choice

Treat AI coding as a revenue project and you'll spend the next few years in appeals. Treat it as a compliance project and you'll be defending records that were built to be defended.

Timothy M. Wogan, JD, CHC, is a healthcare compliance attorney and founder of Wogan Health Law — Counsel for Healthcare. He advises healthcare organizations on compliance, privacy and incident response, and writes the daily brief at Our Office Next Door.

Every morning: the compliance developments that affect your organization, and what to do about them. Start free, no card needed, at ourofficenextdoor.com