The pain woke Sophie up at 2 a.m. Urgent care was open; her dentist was not. So she paid roughly a $100 copay, sat in a waiting room in the middle of the night, and was eventually seen by a clinician in a Hawaiian shirt who declined to prescribe anything stronger than over-the-counter pain relief. His line was that two extra-strength Tylenol and two Advil would do more for her than codeine. The actual problem was a tooth that needed an emergency root canal. The dentist she saw later that day wrote her a prescription for real pain medication, and she never filled it.
So this isn’t a drug-seeking story. It’s a wrong-door story. And the annoying part is that the Hawaiian-shirt doctor was probably right about the drugs. Current ADA guidance for acute dental pain is more or less exactly what he told her — ibuprofen plus acetaminophen outperforms most opioid regimens. The opioid suspicion Sophie sensed was a clinician following a protocol shaped by a decade of overdose deaths. The pharmacology was right. What failed was routing. She was in the wrong building, talking to the wrong professional, paying for a visit that could not have solved her problem even if everyone there did their job perfectly.
Call it the wrong-door visit. American primary care and urgent care have functioned, for years, as expensive triage with a copay. The system rations attention through friction: copays, appointment delays, uncertainty, and the sheer difficulty of figuring out which door to walk through. A dental-related emergency-department visit happens in the US roughly every fifteen seconds. Sophie paid $100 to be told which professional she actually needed.
Seven years ago I interviewed a designer who had built a prototype: a small room with a large chair you sat in, and the room analyzed your health and gave you advice without a clinician anywhere in the loop. I thought it was ridiculous. Medicine wasn’t a vending machine and never would be.
I was wrong, and what’s interesting is what I was wrong about. I had assumed the cognitive part of medicine — pattern recognition, recall, integration of context — was the sacred and irreducible thing. It turns out a lot of that is the most automatable part. The irreducible thing was somewhere else: thresholds, touch, prescription authority, and somebody who can be sued.
I watched the same mistake play out in software. When I was at Facebook, the bottleneck was code, so the entire industry organized itself around finding people who could produce it — the Coding Machine archetype and everything downstream of it: the LeetCode pipeline, the bootcamp factory farm, the whole hiring apparatus tuned for output. Then AI made code cheap, and the bottleneck slid to judgment, taste, ownership, and knowing what was actually worth building.
Medicine is running the same play. The old system used inconvenience and ignorance to suppress demand it couldn’t handle, and AI is dissolving the inconvenience faster than anyone is building the discipline to replace it.
So ask the practical question. If Sophie had typed her symptoms into ChatGPT before getting in the car, would it have known enough to change her first move?
Probably yes. It wouldn’t have handed her a clean diagnosis — atypical presentations are still where these tools fall apart — but it would have pointed at tooth, severe pain, emergency dentist, and what to take overnight. The $100 copay and the lost night were the price Sophie paid for not knowing which kind of professional she needed. That price is the first thing AI takes off the table.
What leaves the visit first is coordination. Pre-visit symptom histories, prioritization, asynchronous messaging, structured intake — none of it is exotic, and none of it requires a clinician’s attention to happen. By the time you sit down in the chair, most of the basic organizing thinking has already happened somewhere else.
Then interpretation starts to go. My own primary-care visits already feel different. The conversation is mostly about which tests to run, and both of us — the doctor and I — are showing up with AI-shaped opinions. When results come back, they’re less interesting as a list of out-of-range flags than as a pattern. A Mayo Clinic analysis of more than two billion lab measurements found that more than half of values shift meaningfully when you account for personal characteristics; the population reference range, the thing that has driven Western lab interpretation for half a century, is closer to obsolete than most patients realize. The trick of noticing that three high-normal values together suggest something a single out-of-range value would not — what a great internist used to do by intuition over years of practice — is increasingly something a model does in the time it takes the page to load.
Which brings me to the money. I pay around $100 to get on a follow-up call to discuss test results. I have paid thousands for panels of tests I am not sure I needed. As long as interpretation was scarce, those prices made some kind of grim sense.
A JAMA Internal Medicine study of answers to patient questions found evaluators preferred AI-generated answers to physician answers nearly four times out of five and rated them more empathetic by almost an order of magnitude. That isn’t a story about doctors being bad. It’s a story about what patients have been buying for decades: basic explanation, professional routing, and the privilege of not having to know which door comes next.
The deeper target isn’t doctors charging money. It’s the billing model. A system that prices professional time around basic explanation gets pressured the moment basic explanation stops being scarce. And the early signs are not encouraging: when federal information-sharing rules produced a 157% surge in patient-portal messages, major health systems didn’t absorb that as overhead. They put billing codes around it and started charging $10 to $50 for messages that require medical decision-making.
Now the harder objection. I had to hear this one from my own primary-care provider before I took it seriously. She told me, more or less, that AI-informed patients are already a problem. They walk in asking for obscure tests meant for specialists. If a result lands out of range — and on any twenty-test panel, base-rate math says roughly two-thirds of healthy people will have at least one abnormal flag — she has to do something about it. Often that means a referral. Sometimes a cascade: more tests, more specialists, more anxiety, more cost, and nothing actionable at the end.
She is right. The optimist’s reply is that the next generation of consumer models will price cost and liability into their recommendations and stop over-triaging. Maybe. But the model that under-triages once and gets sued will over-triage forever after, because that is what the incentive selects for.
So come back to Sophie. The same tool that could have saved her $100 and a sleepless night is also the tool that, used a million times in a million bedrooms, will flood her primary-care doctor with cascades that go nowhere. Both happen, and what’s actually moving is the scarce resource.
Cognition is getting cheap, and the bottleneck is sliding toward everything cognition cannot do by itself: phlebotomy, scans, procedures, prescription authority, specialist time, malpractice liability. US malpractice doctrine still does not shift liability from the clinician to the model vendor, and most FDA-authorized AI medical devices reach the market through the 510(k) substantial-equivalence pathway rather than fresh clinical trials. Those aren’t footnotes. They are the reason this transition will be slow and ugly, and they are also the reason doctors don’t disappear. Somebody still has to sign the order and be sue-able.
The doctor’s office isn’t dying. It’s being unbundled.
My local Whole Foods used to be a place I went to shop. It is now substantially an Amazon grocery-fulfillment node. The doctor’s office is heading the same direction: less the place where all the thinking supposedly happens, more a 24/7 local hub for blood draws, imaging, prescription pickup, body scans, vaccinations, and the procedures that genuinely require a trained pair of hands and a regulated room. The interpretation, the triage, the pre-visit history, the explanation of results, the lifestyle nudges — most of it moves into the cloud and into our pockets. What stays in the room is whatever the body and the legal code physically require to be there.
There is a generational cost in this that nobody is pricing. Primary care wasn’t only routing. It was also the training ground where physicians built the pattern recognition that becomes specialist judgment over twenty years. The boring visit — the sore throat, the blood pressure check, the unremarkable rash — was where the reps happened.
Unbundle it into a fulfillment node and a chatbot, and you get exactly the problem software is now living through with junior engineers: cheap cognition at the front of the funnel, no obvious place for new people to develop taste at the back. I see it already in the engineers I mentor — people who can ship volume but can’t tell you what is worth shipping, because the low-stakes reps that used to build that judgment have been automated out from under them. Medicine is about to run the same experiment with people whose mistakes carry more weight than a bad commit.
So back to Sophie one more time. Hers is the clean case for cheap routing. The system charged her a $100 copay and a sleepless night for not knowing which kind of professional she needed, and a model in her pocket would, more often than not, have walked her past the urgent-care door and toward the dentist.
That’s the easy half.
The harder half is what happens when there are millions of her, each carrying a model that knows enough to do the same thing. The friction that was quietly suppressing demand stops working. The parts of the system that can’t be automated — the blood draws, the prescription pads, the signatures, the liability — get hit with traffic the old building was never designed for.
Stop arguing about whether AI replaces doctors. It doesn’t. It reprices them.
Build for the building we actually need: cheap cognition in the phone, expensive hands and signatures in the room, and an honest plan for where the next generation of physicians is supposed to learn their craft.
Sophie shouldn’t have been in that waiting room at 2 a.m., and the next Sophie won’t be. That’s the easy part. The rest is on us.