On July 1, a health-insurance rule quietly took effect in Indiana, and almost nobody outside billing offices noticed. Insurers can no longer use artificial intelligence as the sole basis to downcode a claim. A human has to review the actual medical record before the payment is cut, and the plan has to tell the provider when AI drove the decision.
Read that twice, because of what it admits was already happening.
Downcoding is the quiet cousin of a denial. The visit gets paid, just at a lower level than the physician documented and billed. A level-4 encounter becomes a level-3. The dollars per claim are small. The scale is not. Run that reclassification across millions of claims through a model that never opens the chart, and you have moved real money out of clinical practices, one imperceptible notch at a time. The physician's options were to absorb the loss or to spend unpaid hours appealing a decision no human had made.
Indiana did two specific things, and both matter more than the headline. It required a human in the loop before an AI downcode stands. And it required disclosure, so the provider knows the machine was involved at all. For years the second point was the real problem: you cannot appeal a decision you did not know was automated, and you cannot hold a black box accountable when you cannot even see it in the room.
This is the part I want every health system and every payer leader to sit with. The debate about clinical AI has been stuck on the wrong question. We keep asking whether the model is accurate. Accuracy is table stakes. The question that actually governs risk is who is named accountable when the model is wrong. Indiana answered it on the payer side by naming a human reviewer and a disclosure obligation. Most states have not, and most health systems have not answered the same question for the clinical AI running inside their own walls.
Because the downcoding story is not really about payers. It is a preview. The same structural gap runs straight through clinical AI at the bedside. When a model flags a case, drafts a note, clears a prior authorization, or suggests a diagnosis, the accountability does not disappear. It gets distributed, and distributed accountability has a way of becoming nobody's. The vendor owns the model. The plan or the system owns the policy. Someone owns the patient in the room. If the contract does not say which is which, those three facts quietly come apart at exactly the moment something goes wrong.
"The debate has been stuck on whether the model is accurate. Accuracy is table stakes. The question that governs risk is who is named accountable when the model is wrong."
There is a second front here that is getting almost no attention, and it is the one I would watch next. While states argue about AI in claims, the pipes underneath American healthcare just changed scale. The federal government's TEFCA framework went from roughly 10 million health-record exchanges to more than a billion in under a year, and HHS just funded auditing to check who is actually following the rules. The interoperability fight, the ten-year war to make records move, is effectively over. The records move now. The next fight is governance of what flows through those pipes, and whether the AI reading a billion freely moving records is held to any standard at all. Indiana regulated one narrow use of AI on one type of decision. The volume of data now available to every model just went up by two orders of magnitude.
So what should a board or an operating team actually do with this. Three things, and none of them require waiting for your state to pass a law.
Indiana did not solve the problem. It named it, on one decision, in one state. The work now is to name it everywhere it already lives, before the next enforcement action or the next patient makes the introduction for you.
Surgery-trained, currently practicing internal medicine (charity care). Advisory practice focused on clinical AI governance, vendor evaluation, and implementation strategy for health systems and health-tech companies. If you are on a board, an operating team, or a clinical-AI committee trying to draw that line, that is the work I do.
Weekly clinical reality on AI, governance, and what actually works inside health systems.