The first AI tool that actually works in healthcare is making medical bills bigger. And everyone agrees.

Caroline Pearson runs the Peterson Health Technology Institute. Earlier this year she held a closed-door roundtable with investors, health plans, and providers to talk about ambient AI scribes. The part that leaked out yesterday is the part I cannot stop thinking about.

Her words: “The investors, the health plans, and the providers, in private, were like, OK, well, it’s quite clear scribes are increasing coding intensity. One hundred percent.”

One hundred percent. Nobody in that room dissented. Nobody tried to argue otherwise. Insurers and providers disagree about almost everything that touches money, and these are the two sides that have spent a decade in trench warfare over claim denials. They agreed. In private, with the recorder off, with nothing to win from the admission, they agreed.

The public version of the story is still the usual fight. On Q1 earnings calls, insurers play the white knight and warn about AI-fueled billing inflation. Health systems push back and claim that AI scribes are finally letting doctors capture the codes they always should have been paid for. Both statements are technically true and both are beside the point. The point, according to the people writing the checks, is that ambient AI is the first productivity win in healthcare at scale, and the productivity is landing on the revenue line, not the cost line.

I have been writing about AI unit economics for a year now. OpenAI wants a robot tax. ChatGPT started selling ads. Neuro-symbolic research keeps showing that efficient beats enormous. Every single one of those stories was about AI meeting its own bill. This one is different. This one is about AI working perfectly and the bill still going up. That deserves a real look.

The technology works. That is the entire problem.

If you have not used an ambient AI scribe, here is what happens. A doctor hits record at the start of the visit. The scribe listens. It transcribes, redacts, structures, codes, and writes the note. By the time the patient walks out, there is a draft SOAP note in the EHR, a list of billing codes suggested, and a to-do list for the follow-up. The doctor reviews, edits, signs. Total touch time has dropped by roughly half in every serious study I have read, and that matches what physicians tell me when I ask them unprompted.

A KLAS Research report from this year found that 93 percent of health systems are projecting moderate to deep adoption of ambient AI tools within the next six months. An MGMA survey of physician practice leaders put current usage at 71 percent. Kaiser, Mayo, Duke, and roughly every other household-name system have deployed at scale. When something reaches 93 percent adoption expectations in a market that normally resists change like healthcare does, you are not looking at a product. You are looking at infrastructure.

The market caught up fast. Microsoft Nuance DAX holds about 33 percent share and is the default for large IDNs. Abridge has raced to 30 percent and a $5.3 billion valuation on $773 million raised, including a $300 million Series E last summer. Ambience Healthcare is at roughly 13 percent and quietly became a unicorn. Suki sits at about 10 percent with $168 million total funding. And Epic, which owns the underlying EHR most of these systems use, just launched a native scribe of its own, which is the kind of move that typically eats half the market within 18 months because distribution beats product in verticalized healthcare. Every one of these companies charges somewhere between $250 and $600 per doctor per month.

So the tools are real, the adoption is real, the money is real, and the productivity is real. Doctors save time. Charts are better. Burnout drops. Everyone I talk to who actually uses these things says the same thing: they would not go back.

And then the bill arrives.

Coding intensity is the rule of the game

Here is the mechanic that most non-healthcare founders miss. In the US, physicians are not paid for what they do in a visit. They are paid for what they document. That is not a cynical framing. It is the actual rule of the game. Medicare, Medicaid, and private payers all reimburse based on Evaluation and Management codes (E/M codes) and, in Medicare Advantage, on Hierarchical Condition Categories (HCCs) that roll up into risk scores. Those codes are built from documentation. Documented more thoroughly, they move up a level. Documented less thoroughly, they move down.

For the last 20 years, the bottleneck on documentation was the physician. Nobody wants to type a nine-point review of systems for a 12-minute visit, so doctors coded defensively and left money on the table. Analysts have been writing about this undercoding for a decade. It was called the “note debt” of American primary care. A huge share of it was real.

Ambient AI walks straight through that bottleneck. It captures every comorbidity mentioned in passing. It pulls forward every ROS element the doctor actually asked about. It documents the time spent on coordination-of-care, which on its own can bump a visit from a level three to a level four. In Medicare Advantage, it surfaces HCCs that would have been missed, which increases the risk score and with it the capitated payment. None of this is fraud. None of this is even aggressive. It is the system working exactly as designed, except the part of the system that used to bottleneck the billing is gone.

A one-level E/M shift, from 99213 to 99214, is worth roughly $50 to $100 per visit depending on geography and payer. A primary care doctor sees 20 to 25 patients a day. Even a modest 20 percent shift in level distribution adds up to six figures per provider per year. Multiply by every PCP in a health system. Multiply that by 93 percent adoption. The math is not subtle.

This is what “coding intensity” means. Not fraud. Not up-coding. Just documentation catching up to the actual work, at the exact moment insurers were starting to squeeze on every other lever they had.

The productivity paradox I did not expect

I spent most of last year writing about AI deflation. Chip efficiency, inference cost collapses, open-source models chasing closed ones, neuro-symbolic approaches cutting training by 100x. The consensus among builders was that AI would do to software margins what commodity hardware did to PC margins: grind them down, slowly and forever. That thesis still holds in most markets.

Healthcare broke it. And I think I know why.

In competitive markets, productivity gains show up as lower prices because a competitor will cut price to win share. The provider captures the gain for a while and then loses it to the customer. Drug prices work this way when generics enter. Cloud prices worked this way for a decade. Consumer software works this way whenever a new entrant has nothing to lose.

Healthcare is not a competitive market in any sense a normal economist would recognize. Patients do not price shop. Insurers do not actually set prices; they negotiate them. Providers are reimbursed by codes, not by outcomes. The entire payment system is built to compensate for documented effort, and “documented” is doing all the work in that sentence. When AI makes documentation effortless, the payment system does exactly what you would expect: it pays more, because more things got documented.

This is the productivity paradox in its purest form. The technology works. The humans are better off. The system gets more expensive. No fraud, no villains, no obvious fix. Just the rules doing what rules do.

I want to flag something for anyone about to rush to judgment. The doctors are not wrong. They really were undercoding. They really were doing uncompensated work. Every study that tried to measure the gap between documented work and actual work found real holes. A lot of what AI scribes capture is work that was genuinely happening and genuinely not getting paid for. If you think primary care is underpaid in the United States (I do), then you think this correction is overdue. The fact that it hits the system as inflation is a feature of our insurance design, not a bug in AI.

The counter-AI is already here

The insurers are not standing still. UnitedHealth, Elevance, Cigna, and Humana have all been building what I would call counter-AI: systems that read the same notes, re-score the same risk adjustments, and downcode claims that look too heavy relative to peer distributions. These are not hypothetical. UnitedHealth has been publicly aggressive about it. The STAT coverage this week flagged insurer algorithms “trying to minimize payments” as the mirror image of provider AI “maximizing codes,” and described the result as an AI coding arms race. That is exactly what it is.

I looked at one of the downcoding tools in a briefing earlier this year. It works the way you would expect. The model sees the structured note, the billed code, and peer distributions for that provider, specialty, and diagnosis. If a particular encounter is two standard deviations heavier than the peer, the system flags it. A lot of those flags turn into partial denials or prepayment reviews. The review itself adds friction, which in billing is currency. The provider eventually accepts a lower level, or burns hours defending the original level, or writes it off. Either way the insurer wins.

This is a classic arms race because neither side can unilaterally de-escalate. If providers stop running ambient AI, they lose the burnout win and return to undercoding. If insurers stop running downcoding AI, they lose a controllable lever on medical loss ratio. The only stable outcome is one where both sides keep running their models, eating the cost of both, and lighting dollars on fire in the middle. And patients, eventually, pay for both sets of compute through premiums and cost-sharing.

What this means for every AI healthcare founder

I think the lesson here is sharper than the usual “regulation is hard in healthcare” warning. It is structural.

In zero-sum billing environments, AI does not sell itself on efficiency. It sells on capture. The vendor that wins is the one who picked the side of the market that captures the new economic value, then built a workflow that side actually wants to adopt. Ambient scribes work because they sell to providers, the side that gets the upside from higher coding intensity. Utilization management AI works because it sells to payers, the side that controls the denial rate. The vendors who tried to split the difference, who pitched “neutral efficiency,” mostly died in pilots. I watched a few of them die in 2024.

So if you are building in healthcare AI right now, stop pitching that your tool “lowers costs for the system.” The system is not the customer. The P&L owner of the line you are affecting is the customer, and the tool has to move that specific P&L in the direction that customer wants. If your tool does not clearly move a P&L, you do not have a product. You have a demo.

There is a corollary that matters for builders in other regulated, zero-sum verticals: insurance claims, real estate commissions, legal discovery, regulatory compliance, even government contracting. In every one of these, “productivity” does not equal “lower prices” automatically. The incumbent billing rules decide who captures the gains. Pick a side before you pick a feature.

What to watch for the rest of 2026

I am watching four things. First, the CMS response. If Medicare concludes that ambient AI is inflating Medicare Advantage risk scores in a way that is not clinically justified, you will see an adjustment to the risk-score model itself, probably in the 2027 rate notice. That is the single biggest reversal risk for Abridge, Ambience, and the rest. It is also the single biggest win for insurers, who have been lobbying for it.

Second, Epic. Epic launching a native scribe inside the EHR is the kind of distribution play that eats VC-funded vendors in 12 to 18 months. If you are a standalone scribe, your moat now has to be something Epic cannot copy in its own roadmap. That is a short list. It is not obvious to me that any independent ambient AI company has it.

Third, the insurer counter-AI. I want to see a public head-to-head: same encounter, same documentation, provider AI codes it, payer AI downcodes it, real dollar impact disclosed. We do not have that yet because neither side wants to release it. The first good leak on this will be the most important story in healthcare AI for the year.

Fourth, the vulnerable providers. Rural clinics, independent primary care, and community health centers cannot afford the $500-per-doctor-per-month scribes, and they cannot afford to ignore them either. In a world where coding intensity is the game and the game is gated by a SaaS subscription, the providers who cannot pay the tool vendor are the ones who get squeezed from both sides. That is where the human cost of the arms race is going to land first, and it will show up in closures and M&A before it shows up in headlines.

One last thought on the broader AI thesis

I keep coming back to the gap between what AI does and what the system around it pays for. Every serious AI productivity story right now has this shape. ChatGPT made content creation cheap, and Google is paying publishers anyway because attention moved. OpenAI made code cheap, and the developer market is spending more on AI than it ever spent on tooling. Ambient scribes made documentation cheap, and bills went up. The technology is doing the deflationary thing you would expect. The economic system is routing the savings somewhere unexpected, and in most of these cases it is not ending up in the hands of the end consumer.

That is the thing I want builders to internalize. The efficiency is real. The deflation is not automatic. The only way to know where the surplus lands is to look at the contract structure of the market you are entering. In healthcare the contract is “paid for what you document.” AI documented more, the contract paid more, and here we are. If you are building into a similar contract in another vertical, figure out where your efficiency is going to land before you write the pitch deck, not after.

I would love to tell you there is a clean fix for this. There is not. Insurers will keep running their models. Providers will keep running theirs. Epic will absorb part of the market. CMS will eventually adjust the risk model. Patients will pay a little more. Builders who picked a side will win the vendor contracts. Builders who did not will call themselves neutral and die in pilots. And somewhere in the middle, doctors will keep using the scribes because the scribes work, and nobody who has gotten an hour of their evening back wants to give it up.

That is the whole story. The first AI tool that truly works in American healthcare is the one that made everything more expensive. Nobody is happy about it. Nobody agrees what to do about it. And in private, at least, everybody agrees it is happening. One hundred percent.

Frequently asked questions

Are AI scribes actually raising healthcare costs in 2026?

Yes. At a 2026 Peterson Health Technology Institute roundtable, investors, health plans, and providers all privately acknowledged that ambient AI scribes are increasing coding intensity. PHTI executive director Caroline Pearson described the consensus as “one hundred percent.” The public dispute is over what to do about it, not whether it is happening.

How exactly do AI scribes drive bills up?

Ambient scribes transcribe entire encounters and surface billable details doctors used to miss, including comorbidities (HCCs) and higher-level Evaluation and Management codes. A one-level E/M shift from 99213 to 99214 is worth roughly $50 to $100 per visit. Across a typical PCP panel of 20 to 25 visits a day, modest shifts add up to six figures per provider per year.

Which AI scribe companies lead the market?

Microsoft Nuance DAX leads at about 33 percent share. Abridge follows at roughly 30 percent on a $5.3 billion valuation and $773 million raised. Ambience Healthcare holds around 13 percent. Suki sits at 10 percent. Epic just launched a native scribe inside its EHR, which is expected to reshape market share over the next 12 to 18 months.

Why is this called the AI coding arms race?

Because both sides now run AI against each other. Providers use ambient scribes and autonomous coders to maximize billable codes. Insurers run their own downcoding and risk-score recalibration models to minimize payments. The result is two AI systems litigating every visit with equal and opposite incentives. It is zero-sum by design.

Isn’t AI supposed to make everything cheaper?

In competitive markets, yes. Healthcare billing is not competitive in the economic sense. Providers are reimbursed for what they document, not for what they deliver. When AI makes documentation effortless, payments rise because more things get documented. The productivity gain lands on the revenue line, not the cost line. That is the paradox.

What should AI healthcare founders learn from this?

Stop selling “neutral efficiency.” In zero-sum billing markets, AI wins on capture, not on cost reduction. Pick the side that owns the P&L your product affects and build for that side’s workflow. Ambient scribes won by selling to providers, who capture the revenue gain. Utilization management AI wins by selling to payers, who control denials. Vendors who tried to serve both died in pilots.