The Toll Booth
The reasoning is getting better every year. The only question left is whether we will be allowed to plug it in.
INSIDE THE LOOP · SUPPLEMENT TO ISSUE 07
The laboratory never loses. The models get sharper, the reasoning improves, the science marches. Implementation is where good ideas go to die, and lately the cause of death is written on an invoice.
In Issue 07 I argued that the problem in patient safety is not the people and not the data, it is a design that waits for the next scheduled look. This is the companion piece. Suppose the design problem is solved and the reasoning works. What actually stops it from reaching a patient? I went looking for the honest answer, and while I was looking, a German hospital-IT executive published the same argument in the same week, from four thousand miles away, about a health system nothing like ours [7]. When two people who have never met describe the same wall from opposite sides of an ocean, it is worth asking what the wall is made of.
What the machine can already do, and what it cannot yet
Be honest about what the reasoning can and cannot do. The honesty is the whole credibility.
Start with a claim I can defend in both directions. Machine reasoning is improving quickly, and it is not a doctor in a box. Both are true, and anyone who tells you only one of them is selling something. On structured medical knowledge, frontier models now score in the low nineties on licensing-exam reasoning, better than most people expect [9]. On the hardest open-ended diagnostic cases, the same models still fall to around half right, which is nowhere near good enough to trust alone [10]. So the sober read is this: the technology is already strong at bounded, well-posed questions, still weak at open-ended ones, and getting measurably better at both every few months.
That is exactly why my own demonstration does not ask the model to diagnose. It asks a narrower and more answerable question. Given this patient’s trajectory, the orders already written, and the evidence around them, is this permitted action the right action right now. That is a bounded problem the reasoning handles well today, and it is the kind of problem that improves fastest as the models improve. I am not claiming a breakthrough in judgment. I am claiming that a capability which did not exist five years ago now exists, it is cheap, and almost nobody is testing it against live clinical logic. We are among the first who actually are, on real public data rather than a slide.
Two prices, one of them quietly enormous
The thinking is nearly free. The permission to think is not.
So if the science is improving, what is the wall? Money, but not the kind you would guess. I priced both ends of this in a companion research note. A single reasoning pass over a patient, the actual thinking, costs somewhere between one and ten cents depending on the model, and less once the patient’s context is cached and reused [1]. The other end, the plumbing that lets that reasoning see a real patient and act, is where the zeros pile up. A read-only feed from a major electronic health record starts around fifteen thousand dollars. A live, two-way interface that can take in vitals and send back an order runs one hundred fifty thousand to three hundred thousand dollars and most of a year to build [2]. A mid-sized system running fifty such interfaces reported more than six hundred thousand dollars a year just to keep them alive [3]. The thinking is nearly free. The permission to think is not.
A price, or a toll
There is a line between charging for work and charging for passage. Two countries crossed it the same week.
Here is where the German commentary matters, because it is not my voice and it is not my market. Bastian Stockhausen, who runs a hospital-IT services company in Germany, called the mechanism by its financial name: Cause of Death, EBIT [7]. His argument is the one I have been circling. An interface is a bottleneck through which all the information must pass, and large incumbent vendors price that passage far above what it costs to build or maintain, not for technical reasons but strategic ones. He gives a real example: a vendor asked for more money to supply a data server by itself than it had quoted for that same server bundled with its own patient portal, because the hospital had chosen a competitor’s portal. The partial service cost more than the whole package. As he puts it, that price has nothing to do with cost. It is the exercise of market power to punish a free purchasing decision.
His analogy is better than mine, so I will borrow it. Roughly a fifth of the world’s oil passes through the Strait of Hormuz, a channel a few kilometers wide. Whoever controls that passage does not need to own the oil. Controlling the transit is enough to move prices, and the bill is never paid by the powerful at the bottleneck, it is paid by the consumer at the pump. An interface in a hospital is the same shape. What is a consumer at the pump is a patient at the bedside. And both Stockhausen and I are careful about the same line, because it is the line that keeps this honest: genuine integration work deserves a real profit, and no serious person argues otherwise. The toll is not profit. The toll is the part of the price that exists to exclude a competitor rather than to deliver a service.
Even the law gets gamed
Germany made open interfaces mandatory. Guess what happened to the price.
You might assume a legal mandate for open interfaces settles this. Germany tried exactly that. Under section 373 of its social code, the agency Gematik was directed to define binding, standardized, modern interfaces for hospital systems, and compliance is mandatory for manufacturers [8]. Open data exchange there is not a favor. It is the law. And here is the sentence Stockhausen reports hearing from inside the industry, which I cannot improve on: if we are legally obliged to do so, we will just make it even pricier [7]. The mandate was not read as an instruction to open up. It was read as a new place to collect at the bottleneck.
We have the American version of the same lesson. Our Cures Act rules require certified records to expose a standardized, modern interface without special effort, and they bar fees designed to shut out competing apps [5]. A national exchange framework is live and expanding [6]. That genuinely lowers the cost of reading standardized data, and it is real progress. But it does not yet deliver a free, real-time, write-capable production feed, so the last mile stays tolled. Two continents, two mandates, one workaround: when you cannot refuse to open the gate, you raise the price of walking through it.
The four-billion-dollar anchor
A sunk cost that large stops being an investment and starts being an instinct.
One more number, because it reframes the incentive. Kaiser Permanente spent an estimated four billion dollars standing up its electronic record platform, at the time the largest privately funded record in the world [4]. I am not mocking that figure. It bought real capability, and running a system at that scale is genuinely hard. I raise it because a sunk cost that large changes behavior. It creates a powerful instinct to protect the investment, to keep spending to keep the legacy alive, and to treat every new connection as a toll that helps recover what was spent. That instinct is human and it is understandable. It is also, step by step, how an industry talks itself into spending more and more each year to stand exactly still.
Meanwhile, the patient walked out of the building
While hospitals argue over the price of a feed, the data got up and left.
Now the part the toll collectors are not pricing in. The continuous physiologic stream that hospitals guard behind expensive feeds is increasingly being generated somewhere else entirely: on the patient, at home, on hardware they bought themselves. In 2024 the Food and Drug Administration cleared the first over-the-counter continuous glucose monitors, Dexcom’s Stelo and Abbott’s Lingo, sold without a prescription for roughly the price of a phone case a month [11]. The wearable medical device market was about fifty-four billion dollars in 2025 and is growing at double-digit rates, with home health already its largest segment [12]. Remote patient monitoring is a market in the tens of billions and roughly doubling across the decade [12].
Read that against the interface problem and the picture turns. A reasoning layer runs over a consumer data stream for pennies, and that stream does not sit behind a forty-thousand-dollar gate, because the patient owns it. If the walled systems keep the drawbridge up, the data gravity and the innovation do not wait politely at the moat. They move to the wrist. The real risk to legacy healthcare is not dramatic disruption in a boardroom. It is quieter and more humiliating: getting routed around while spending more every year to defend a gate the world is learning to walk past. The reasoning that a hospital cannot afford to connect to its own record will run, cheaply, on the monitor a patient bought on the internet. That should terrify the toll collectors, and it should encourage the rest of us.
What this is actually asking for
This is not a demand to work for free. It is a demand to stop paying at a booth that produces nothing.
So let me be plain about the goal of this piece, because it is not merely to describe the weather. It is a call to stop treating the toll as a fact of nature. The ask is narrow and it is fair. Price an interface like the service it is, not like a checkpoint. Honor the open-standard mandates in their spirit and not just their letter. And let an integrated system, or any willing partner, test a promising method on data it already owns without a six-figure tax at the door. None of that asks a vendor to abandon profit. Stockhausen and I agree on that completely. It asks only that the price reflect work performed rather than passage permitted.
Bottlenecks do not fall because the gatekeeper has a change of heart. They fall when enough people downstream, the buyers, the clinicians, the hospital IT leaders, and yes the readers of a newsletter like this one, refuse to keep pretending the toll is a cost. Say the quiet part out loud often enough and charging the toll becomes an embarrassment. That is the grassroots version of progress, and it is the version available to people like us who have no boardroom and no leverage except being right and being loud.
It is not only greed, and the regulator is not the enemy
Hand the skeptic their due, and the argument gets stronger, not weaker.
I would be a poor journalist if I let greed carry the whole blame, so here is the fair accounting, and it holds a surprise. Money is not the only thing slowing this down. Someone has to own the liability when a recommendation is wrong, clinicians have to trust it, and governance committees have to approve it. All real. But the regulatory picture is lighter than most people assume, and it just got lighter. Under the 21st Century Cures Act, clinical decision support that lets a clinician independently review the basis of its advice, and does not ask the clinician to rely on it primarily, is generally not regulated as a medical device [13]. In January 2026 the Food and Drug Administration revised its guidance to cut red tape further, extending enforcement discretion even to tools that offer a single recommendation when only one is clinically appropriate, while leaning harder on one thing in particular: transparency about the data, the logic, and how a recommendation was reached, especially for tools driven by artificial intelligence [14].
Sit with that, because it describes the design I already chose. My demonstration is advisory. A physician accepts, adjusts, or rejects at every cycle, and the consult note exists precisely so the basis can be reviewed: the traceable reasons, the supporting evidence, and an honest statement of what the model could not see. Those are not decorations. They are the exact properties that keep transparent, human-in-the-loop support in the lighter-touch lane the FDA just widened. So I will not pretend a regulatory wall is the thing stopping this, because for advisory reasoning it largely is not.
I want to be very clear about tone here, because it matters. I am not anti-regulation and I am not anti-FDA. My proposed method needs oversight, and it should have it. A regulator asks the questions an inventor is too close to see, and proper oversight is not a tax on a good idea, it is part of what makes the idea safe enough to trust. I want the agency in the room, as a partner rather than an adversary. My one honest frustration is timing, not authority. Regulators move at the speed of process, and this technology moves at the speed of a software release. By the time guidance catches up to one generation of a tool, the market has morphed into the next. The answer to that is not less oversight. It is oversight that engages earlier, iterates faster, and treats a builder as a collaborator. The January revision is a sign the agency knows this. I want more of it, sooner.
And where the wall is real, I will say so plainly, because a careful reader deserves the whole map. Two features of this work can pull it back under full device regulation. Reasoning over a continuous signal from a monitor is treated differently, in the agency’s eyes, from reasoning over a lab value a clinician could look up, and my method watches trajectories built from device data. And the moment the loop actually closes, when the reasoning acts instead of advises, it is a device, full stop, and it should be. None of that is a reason to stop. It is a precise map of where the honest review belongs. But notice what none of it is: not one of those obstacles is a reason the plumbing should cost a quarter of a million dollars before the first patient is ever watched.
The cost of doing nothing
Inaction is not the neutral choice. It is the expensive one, and it is not paid in dollars alone.
Every argument in this piece has an alternative, and it is the one we are already living: change nothing. So price that honestly too, because doing nothing is neither free nor safe. If the current course holds, the trajectory is not in dispute. National health spending reached eighteen percent of the American economy in 2024, about five point three trillion dollars, and the government’s own actuaries project it to climb past twenty percent within the decade [15]. That is one dollar in five. The outcomes that money buys have barely moved, the avoidable harm I documented in Issue 07 keeps recurring at rates that have not fallen in a generation, and the population that must be cared for grows older every year, which only raises the demand and the bill. Hold the line exactly where it is and look at what you have chosen: prices that keep rising, outcomes that keep flatlining, and a widening gap between what the system costs and what it delivers. Inaction has a body count and it has an invoice. We simply do not print them on the same page, so we let ourselves believe they are not there.
Where I land
AI reasoning will get better. That is a fact, not a hope. The laboratory will not be stopped. The implementation might be, and if it is, the cause of death will not appear in any journal, because greed does not file a case report. But the way through is already arriving from three directions at once. The reasoning is getting cheaper and stronger. The standards are being pried open, slowly, on two continents, and the regulator has started clearing a path rather than blocking one. And the data is migrating onto devices the patient already owns, where no toll can reach it. The machine can already think well enough to help. The open question is whether we will let it, or spend another decade paying at the booth while the patient, wearing tomorrow’s monitor on her wrist, walks straight past us and out the door.
One last thing, from me
I will end closer to home than I usually allow. I have decided that aimedagent.net is finished, at least as a thing I keep building. It is live, it works, and it proves what I set out to prove. If someone finds a bug, tell me and I will fix it, but the demonstration has done its job. This argument was never going to be won in my code. It will be won, or lost, by whether the rest of us insist on it.
I can afford to be patient about that, and I want to be honest about why. Like a lot of families, mine has quietly taken on more of its own healthcare, because the system we pay so much for makes us wait months for an appointment and then routes that appointment through an approval process built to contain cost rather than to help. That experience changed me. I wear an Oura ring every day and a new Apple Watch, and I do a careful amount of my own AI-assisted research on my health, not to replace my doctors but because I have learned I have to meet them halfway and walk in prepared. I am old enough to know how to work this system. I even pay my primary a monthly fee, out of my own pocket, for nothing more than a better place in line if I need a consult. It is not cheap, and I can manage it.
What I cannot stop thinking about is that my children will manage it less easily, and my children’s children may not even know the system was once meant to work differently. They will not have my inside knowledge, and most of them will not have the money to buy their way up a waitlist. If we do not find a way, or make a way, to lower the barriers this whole piece has been about, they will inherit a thinner, slower, more expensive version of care and be told it is normal. For some of them that difference will be measured in comfort. For a few, it will be measured in years. That is not an inheritance I am willing to hand down without a fight.
So I am going to step away from this project for a while and give some other work its turn. The passion has not dimmed, only redirected, and I will keep publishing here, because the writing may be the part that matters most now. The machine can already think well enough to help. Whether we let it, and whether we finally tear down the toll booth standing between it and the people it could serve, is up to far more of us than me. I have said my piece. Now I am asking you to say yours.
The lab never loses. Let us stop making the toll the reason our children do.
References
[1] Anthropic. Claude model and API pricing, accessed July 2026. Per-pass cost is an estimate from published token rates, detailed in the companion research brief. https://platform.claude.com/docs/en/about-claude/pricing
[2] Arkenea. Epic EHR Integration: costs, APIs, and best practices, 2026. Read-only and bidirectional interface cost ranges. https://arkenea.com/blog/integrating-healthcare-app-with-epic-ehr/
[3] Nirmitee.io. Mirth Connect interface build cost and pricing matrix, 2026. Multi-interface program and annual maintenance figures. https://nirmitee.io/blog/mirth-connect-interface-build-cost-pricing-matrix-2026/
[4] InfoWorld. How Kaiser bet 4 billion dollars on electronic health records, and won. Kaiser Permanente HealthConnect implementation cost. https://www.infoworld.com/article/2272927/how-kaiser-bet-4-billion-on-electronic-health-records-and-won.html
[5] Office of the National Coordinator for Health IT. ONC Cures Act Final Rule: standardized FHIR API and information blocking. https://www.healthit.gov/topic/oncs-cures-act-final-rule
[6] Office of the National Coordinator for Health IT. Trusted Exchange Framework and Common Agreement (TEFCA), status 2025-2026. https://www.healthit.gov/topic/interoperability/policy/trusted-exchange-framework-and-common-agreement-tefca
[7] Stockhausen B. Cause of Death EBIT: Expensive interfaces block healthcare digitalization. heise online, 17 July 2026. Commentary; includes the FHIR-server pricing example, the Strait of Hormuz analogy, and the reported industry quote. https://www.heise.de/-11368331
[8] gematik / ISiK (Informationstechnische Systeme im Krankenhaus), Section 373 SGB V. Binding, FHIR-based hospital interfaces mandated for manufacturers; Basis Stufe 3 deadline July 2025. https://kodjin.com/blog/isik-compliance/
[9] Benchmarking large language models on the United States Medical Licensing Examination for clinical reasoning, 2025. Frontier models reach the low nineties on structured licensing-exam reasoning. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12796295/
[10] DiagnosisArena: Benchmarking Diagnostic Reasoning for Large Language Models, 2025. Top reasoning models reach roughly half accuracy on the hardest open-ended diagnostic cases. https://arxiv.org/abs/2505.14107
[11] MedTech Dive and CNBC, 2024. FDA clears first over-the-counter continuous glucose monitors: Dexcom Stelo (March 2024) and Abbott Lingo (June 2024). https://www.medtechdive.com/news/dexcom-sells-stelo-over-the-counter-cgm/725310/
[12] Fortune Business Insights and Grand View Research, 2025-2026. Wearable medical device market near 54 billion dollars in 2025 with home health the largest segment; remote patient monitoring in the tens of billions and growing at double-digit rates. https://www.fortunebusinessinsights.com/industry-reports/wearable-medical-devices-market-101070
[13] U.S. Food and Drug Administration. Clinical Decision Support Software, guidance for industry; non-device CDS criteria under Section 520(o)(1)(E) of the FD&C Act, added by the 21st Century Cures Act. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software
[14] U.S. Food and Drug Administration, revised Clinical Decision Support Software guidance, January 2026; enforcement discretion extended to single-recommendation tools and heightened transparency expectations for AI-driven CDS. Summary: Covington & Burling, Five Key Takeaways, January 2026. https://www.cov.com/en/news-and-insights/insights/2026/01/5-key-takeaways-from-fdas-revised-clinical-decision-support-cds-software-guidance
[15] Centers for Medicare & Medicaid Services, National Health Expenditure data and 2024-2033 projections; health spending was 18.0 percent of GDP in 2024 (about 5.3 trillion dollars) and is projected to exceed 20 percent within the decade. https://www.cms.gov/data-research/statistics-trends-and-reports/national-health-expenditure-data/nhe-fact-sheet
About the author
Daniel Pettus is the founder of Inside the Loop and the inventor of AI MedAgent, a proof-of-concept closed-loop medication management architecture. He spent forty years in medical device and health IT leadership at Alaris, CareFusion, and BD. He co-founded iMetrikus in 1998 and contributed to IHE Patient Care Device interoperability standards. He holds two patents with a third provisional filed.