I have spent years watching technology try to get into the operating room. Most of it never does. So I noticed something about healthcare AI's big summer: the industry has arrived at the hospital in force, and it has set up everywhere except the room where the hospital makes its money.
The summer was real. ChatGPT Health opened to every American adult. Medicare put AI in front of elective surgeries in six states. And the frontier labs' healthcare teams are all shipping against a version of the same roadmap: prior authorization, clinical documentation, coding, care coordination, credentialing.
It is a good roadmap. The workflows are real, the burden they address is real, and the text-in, text-out shape of the work fits what current models genuinely do well.
But the roadmap has a blind spot, and it is an expensive one. Every workflow on that list has its highest-dollar instance in the same place, and that place does not appear on anyone's roadmap by name: the perioperative suite. The operating rooms, and everything that feeds them.
When I guided fly fishing, every river had the same geography. There were easy pools near the access point, and they were crowded by nine in the morning. The water that held the fish was upstream: harder to reach, harder to read, harder to wade. Value on a river concentrates where getting to it costs something. The healthcare AI roadmap, right now, is an industry fishing the easy pools.
The Economic Engine of the Hospital
Surgical services drive roughly half of a typical US hospital's revenue. The margin story is even more concentrated: an operating room minute costs somewhere between $36 and $46.
And most hospitals leave those minutes empty. Utilization hovers near 60% against a target of around 85%, and a single same-day cancellation forfeits $5,000 to $12,000.
The empty minutes are the most expensive silence in the building.
Vendors and consultancies will happily extend that arithmetic into national estimates that run to the tens of billions a year. I distrust numbers that big on principle, and the case does not need them.
Because here is how enterprise healthcare AI actually gets bought. The labs are selling deployments directly to health systems, and the CFO who signs that contract does not renew it because clinicians enjoy drafting emails with it. They renew it because something moved a number they report to the board. OR throughput is the biggest operational number a hospital has.
Perioperative is not a new vertical for healthcare AI. It is the proof of value that makes the rest of the enterprise deal defensible.
Why the Roadmap Skips the Room
There is a reason the roadmap skips the room, and it is a good one. The operating room means deep integration, physical-world state, and surgeon change management. It is the hardest room in the hospital to ship software into. Better, the argument goes, to let specialist applications own that depth while the platforms stay with shallow, replicable workflows.
That is true of the end state and false of the entry point.
The entry point is not a robot and not a scheduling brain. It is the same text-in, text-out work the roadmap already covers, pointed at its most valuable cases. And text in, text out is not just the shape models are good at. It is the shape where a human still signs before anything happens, so a wrong answer costs a correction instead of a case.
Surgical prior authorization is the highest-denial, highest-dollar subset of a workflow the industry already ships, and this year it stopped being hypothetical: Medicare's own pilot chose elective procedures as its first target. Operative notes and anesthesia records are the messiest documents in the hospital, which is another way of saying they are where documentation AI is worth the most. Pre-op instructions and surgical coding: same shape, bigger checks.
The platforms have already built everything this requires: HIPAA infrastructure, FHIR integration, enterprise contracts. What the entry point needs is knowing which workflows to pick, what the evals must measure, and why hospital AI pilots die: integration debt, clinical trust, workflow disruption. Most of all, the surgeon who was never asked.
From there the path stages naturally. First, text in and text out. Second, systems that read operational state: tools that watch the day's schedule and say, in plain language, what is slipping and why. That is harder than it sounds, because the OR's state changes faster than most systems can read it, and a recommendation built on ten-minute-old data is a new kind of wrong answer.
Third, and only after the trust is earned, coordination across the OR, the recovery unit, sterile processing, and staffing. A model asked to schedule will always produce something plausible. Plausible and feasible are different properties, the difference between a mountain bike line that looks clean from the trailhead and one you can actually ride, and only one of them cancels cases. Getting that right is a genuine capability frontier, which is exactly why the deep end of this pool belongs to frontier labs and not only to application vendors.
That is the version where AI helps surgery happen. The first AI to actually reach the OR door this year came with a different job.
The AI at the Door Is Paid to Say No
That job is deciding whether surgery happens at all. Medicare's WISeR pilot put an AI screen in front of elective procedures in six states, and it arrived with a business model that should worry everyone who wants this technology to succeed: the firms running the reviews are paid a percentage of the savings from care they avert.
Nobody has to instruct a model to over-deny. The incentive does the tuning. Every ambiguous case, every borderline threshold, every insufficient-documentation default drifts toward the side that pays, and the early reporting reads like the incentive predicted: delays for patients and physicians.
Guiding taught me how much the goal shapes the trip. A guide who just needs the client to catch a fish takes them to the stocked pool and has a small one in the net by ten. A guide after the best day the river can give rows past it, trades the sure thing for harder water, and puts the client on a real chance at a big fish. Same river, same skill, two different days, and the difference is what the guide believes they are being paid for.
The same choice is waiting, in quieter form, for every AI product that touches the surgical pipeline. A scheduling model looks smarter if it steers around the specialties whose cases are hardest to predict. An optimization tool hits targets faster by avoiding messy service lines instead of helping them. The tempting tune and the responsible tune point in opposite directions, and the business model decides which one wins. The version worth building uses AI to help hospitals absorb uncertainty, not to make uncertain cases someone else's problem.
Responsible AI in surgery is not a model property. It is a business model.
Whoever builds the perioperative layer of the healthcare AI stack will make that choice early, in the incentive design, long before any model card gets written.
What the Room Demands
I have written before about building AI where being wrong is expensive, and I learned some of it the embarrassing way, by showing off work I had not yet stress-tested. The OR is the room where that kind of lesson stops being cheap.
The workflows there are dense with constraints. The users are experts with no patience for tools that waste their seconds. The cost of a wrong answer is measured in delayed cases and eroded trust, not in mild embarrassment. A team that earns adoption in that room has proven something no demo can prove, and the discipline it builds travels to every easier room in the building.
Ambient scribes spent a decade earning their way into the clinic. The same curve now bends toward the room where the hospital makes its money. The roadmap that arrived this summer is right about the work. It has just not named the place where that work is worth the most and costs the most to get wrong.
Somebody will build the perioperative layer of the healthcare AI stack. The open questions are whether they will build it on incentives a patient would recognize as help, and how much trust gets spent before they do.
The hardest room in the hospital is also the one that pays. Build there first.