The Peterson Health Technology Institute published a report in April that everyone in health care cited and nobody seemed to actually read.

The headline finding got picked up widely: AI deployed on prior authorization and medical billing is not reducing system costs, and is probably increasing them. Providers automate submissions, payers automate denials, and the volume of transactions goes up while the underlying dispute stays exactly where it was. Participants at PHTI's workshop had a name for it. Bot wars.

That framing is now everywhere, usually accompanied by a rueful observation that this is all terribly wasteful. Which it is. But calling it an arms race, or waste, quietly implies that it is nobody's fault, that the participants are trapped, and that if only everyone would stand down we could go back to something sensible.

Look at the numbers underneath and a different picture appears. Nobody is trapped. Everyone is behaving rationally. The system is working exactly as designed, and it is producing precisely what it is being paid to produce.

The tell is in the scribe data

Start with ambient AI scribes, the least controversial application in the whole category. They listen to the patient encounter and draft the note. Doctors like them. They are marketed as a cure for burnout.

Here is what the evidence shows. Time savings from an AI scribe come to roughly two hours per physician per year. Not per week. Per year.

Here is what else the evidence shows. At one multihospital system, after deploying an AI scribe, level 5 encounters for established patients rose 5% and level 4 encounters rose 7%, producing an average revenue increase of about $1,004 per provider per month.

Two hours a year of relief. Twelve thousand dollars a year of revenue.

Now ask yourself which of those two numbers is on the slide when the vendor pitches the health system CFO, and which one is in the press release. The product is sold as a burnout intervention. It functions as a revenue instrument. That is not a scandal, and the coding may even be accurate, in the sense that the complexity was always there and simply was not being documented. But accuracy is not the same as neutrality. A tool that finds every billable complexity and never finds a reason to bill less is not a documentation tool. It is a ratchet.

PHTI is blunt about the consequence. Provider deployment of AI is increasing billing intensity and inflating medical spending. One study cited in the reporting put the additional spending from AI-enabled billing intensity at $2.3 billion for the settings examined.

The payer response is not a defense, it is a tax

Insurers noticed. Their answer, per PHTI, is across-the-board downcoding and other broad reimbursement reductions, often executed by their own AI, which flags high-complexity evaluation and management codes that look inconsistent with the documentation.

Read that carefully, because it is the most important sentence in the report and it is not about technology at all.

Payers are not adjudicating each claim more accurately. They are applying a blanket cut in response to an aggregate trend. That is a tax on the pool. And a tax on the pool falls hardest on whoever is not doing the thing that triggered it.

PHTI says so explicitly. The reimbursement cuts may disproportionately harm providers who have not adopted AI tools. Independent practices. Rural hospitals. Safety-net providers. The ones who cannot afford the scribe, cannot afford the coding engine, and cannot afford the consultants who tune them.

So the sequence is: large systems buy AI, their coding intensity rises, payers cut rates across the board, and the cut lands on everyone including the clinics that never bought anything. The arms race does not merely waste money. It transfers it, from the small to the large, and from both to the vendors selling weapons to each side.

That is the part the "waste" framing obscures. Waste implies the money evaporates. It does not evaporate. It moves.

The friction was never a bug

There is a comforting story in which prior authorization is an ugly but well-intentioned mechanism that has become inefficient, and AI, properly deployed, could streamline it.

The data does not support the comfort. Across four major insurers there are more than 5,000 procedure codes requiring prior authorization, and only 3% of those codes require it across all four. Ninety-seven percent of the friction is not clinical judgment. It is variation for variation's sake. Each plan maintains its own definitions of medical necessity, its own documentation requirements, its own procedural hoops.

Complexity of that kind is not an accident that accumulated. It is a moat. Every incompatible rule is a small cost imposed on someone else's balance sheet, and in a system where prior authorization is estimated to suppress as much as 5% of medical and drug spending, friction is a revenue strategy that happens to look like paperwork.

Which is why layering AI on top does nothing. As one provider told PHTI's workshop, bots do not get tired of asking questions, so the review queue keeps growing. You have automated a process whose function was to be costly. Making it cheaper per transaction just means more transactions.

Each prior authorization cycle costs a plan $40 to $50 and a provider $20 to $30. Halve those costs and double the volume and you have accomplished nothing except making the fight faster.

Nobody in the room is misbehaving

This is the uncomfortable conclusion, and PHTI's Caroline Pearson states it plainly: under current incentive structures, no stakeholder is positioned to deploy AI in ways that reduce friction across the system. Efficiency gains accrue inside organizations. They do not flow through to lower costs.

A hospital CFO who declines to buy the coding AI while competitors buy it is not taking a principled stand. They are accepting a permanent revenue disadvantage against a downcoding regime calibrated to peers who did buy it. An insurer that declines to automate review while providers automate submission is choosing to pay claims it could have denied. Each actor is doing the correct thing given what the others are doing. The equilibrium is terrible and nobody can unilaterally exit it.

Of the estimated $350 billion in annual U.S. administrative waste, PHTI attributes $266 billion to administrative complexity. That figure predates the AI, and the AI is being sold as the solution to it. What the AI is actually doing is metabolizing it faster.

What would actually work, and why it will not happen soon

PHTI's participants identified the lever, and it is not technological. Reimbursement policy is the strongest tool for driving systemwide savings. Everything else is a faster horse.

Some structural fixes are in motion. CMS-0057-F requires Medicare Advantage, Medicaid and marketplace plans to stand up prior authorization APIs by January 1, 2027, with response deadlines of 72 hours for urgent requests and seven days for standard ones. More than 50 health plans have pledged real-time responses to 80% of electronic requests by 2027. Standardized APIs and digitized medical policies would reduce the information asymmetry that the bots are currently fighting over.

But an API standardizes the pipe, not the policy. If the 5,000 codes and the incompatible medical necessity definitions survive, faster plumbing just delivers the same dispute at higher speed. And the more radical options PHTI raises, such as abandoning up-front review in favor of postpayment audit, or letting providers skip prior authorization entirely in exchange for a discounted rate, would require someone to give up leverage. Nobody gives up leverage because a report recommends it.

The only guaranteed winner

Set aside providers and payers for a moment and ask a simpler question. Who is definitely making money on this?

The vendors. The coding engines, the ambient scribes, the prior authorization automation platforms, the payer-side review models, the revenue cycle management firms selling optimization to hospitals and denial management to insurers, sometimes within the same corporate family. That layer collects regardless of who wins any individual claim, and it collects more the longer the war goes on.

PHTI notes there is no evidence that cheaper cost per prior authorization decision translates to a lower average cost per claim once the cost of the AI itself is counted. Read that as what it is. The savings, such as they are, are being consumed by the tools that produced them.

An arms race is a bad metaphor because arms races can be won. This is closer to a protection racket that both sides pay into voluntarily, on the theory that the other side is paying too.

And the patient, who has no AI, no bot, and no seat at the table, receives the invoice.

Further reading