Finally, a Good Use for AI

Most of the AI I see lands somewhere between annoying and insulting. It writes the emails we wanted to write ourselves and paints the pictures artists wanted to paint. It floods the internet with text nobody asked for and nobody will read. And under all of it is the same pitch to somebody's boss: this can do the job for less. The backlash is earned. Point a machine at the work people like doing, and at the paychecks attached to it, and they're going to be angry.

But there's a different kind of work, the kind there will never be enough people to do no matter how the next century goes. That's where AI stops being a toy and starts being the only answer we have.

Healthcare is that work.

The math no one wants to say out loud

By the WHO and World Bank's count, 4.5 billion people, more than half of everyone alive, can't get the essential health services they need. That isn't a statistic about poor countries somewhere far away. It is the median human experience.

The easy assumption is that this is a poverty problem and growth will fix it. It won't. The United States is the richest healthcare market on earth and it's short on doctors too. The Association of American Medical Colleges puts the shortfall at up to 86,000 physicians by 2036. More than 92 million Americans already live in places the federal government has flagged as not having enough primary care doctors. And those numbers assume demand stays suppressed. The AAMC also ran the scenario where people who face barriers to care start using the system at the same rate as everyone else: the country needs more than 200,000 additional physicians, today.

You cannot train your way out of this

The obvious response is to make more doctors. We should. We also can't do it fast enough to matter.

Becoming a physician in the United States takes eleven to sixteen years: four of college, four of medical school, then three to seven of residency before a new doctor can practice on their own. Residency is the real bottleneck, because residency slots are mostly funded by Medicare, and Medicare's cap on them hasn't moved since 1997. The most ambitious fix in Congress adds about 14,000 slots over seven years. Meanwhile demand keeps climbing as the population grows and ages, and nearly a quarter of active physicians are 65 or older themselves.

The next answer is nurse practitioners and physician assistants, and it's a better one. Training takes six to eight years instead of eleven to sixteen, and tens of thousands graduate every year. The catch is that the shortage projections already count them. The gap is what's left over after that pipeline does everything it's supposed to. And an NP is still one person seeing one patient at a time.

More funding would help. So would more residency slots and better distribution. None of it changes the basic math: the pipeline runs on a decade-long clock and the need is here now.

This is the part people skip. You don't solve this by trying harder.

The choice we have been avoiding

That leaves a choice we've spent years pretending we don't have to make. We can accept that healthcare won't work for billions of people, which is the quiet status quo. Or we can let an AI do the clinical work.

I want to be precise about that second option, because the comfortable version of it is a trap.

Why "AI helps the doctor" is not enough

The safe thing to say is that AI will assist doctors and make them more productive. A copilot. A scribe that writes the notes, a tool that drafts the summary. Everyone nods. Nobody feels threatened. And nothing changes, because the human is still the rate limiter.

Do the arithmetic. A doctor has roughly eight working hours in a day. If that doctor has to personally touch every patient interaction, even briefly, you hit a ceiling fast. Faster notes don't move that ceiling. Neither does a better dictation tool.

The bottleneck was never just paperwork. It was clinical attention, and one trained human can only pay attention to one patient at a time.

If you want one doctor to effectively care for ten or twenty times the patients, and the numbers above say you have to, the AI can't be a helper sitting beside the doctor. It has to do the actual clinical work: refill the prescription, order the labs, set up the referral. It diagnoses what it can already diagnose reliably and makes the everyday clinical decisions a competent doctor would make in the same seat. It runs the ordinary case start to finish and pulls in a human only when a case needs one. Over time the line moves, from routine calls to harder ones, and eventually to the decisions no clinician has a clean protocol for. The doctor stops being a faster individual contributor and becomes the supervisor of a clinical workforce that happens to be software. That's the leap.

Say "one doctor will see twenty times the patients" while the doctor is still in the loop on every visit and you're describing something impossible.

Let the AI run the encounter and the same sentence becomes arithmetic.

The word that makes people uncomfortable is "autonomous." But the alternative, a faster assistant, is just a polite way of leaving the bottleneck exactly where it is.

The objections, taken seriously

The honest pushback is obvious. Can we trust an AI to do this? Not all at once, no.

You earn autonomy in steps, the way driving did: cruise control, then lane-keeping, then a system that handles the whole highway, each stage proven before the next is allowed. Medicine starts the same way. There are already head-to-head studies where models match clinicians on routine diagnosis, so that's where the boundary goes first, and it moves when the evidence moves. Nobody serious is proposing to hand the hardest cases to a model tomorrow.

The harder objection is about scale. A million doctors make a million different mistakes. One model makes the same mistake a million times. A blind spot doesn't stay in one exam room, it ships to every encounter at once. All true. But the same property that spreads the error spreads the fix. Catch it once and it's fixed everywhere, that day. You don't wait a generation for the profession to absorb the lesson. And you will catch it, because every encounter is logged. The kind of flaw that hides inside one doctor's practice for thirty years shows up here as a pattern in the data. Medicine has never had a clinician you could fully audit. Now it does. I'd rather have the failure mode you can find.

Then there's the comparison problem.

Critics measure AI against an attentive doctor with unlimited time and perfect judgment. That doctor is fiction. The real comparison is the average doctor on an average Tuesday, who is human and makes mistakes, and for billions of people even that doctor isn't on offer.

The actual alternative is a wait measured in months, a clinic that doesn't call back, or nobody at all.

And yes, I opened with a global number and then spent the whole argument in America. Fair. An AI can't stock a pharmacy or build a lab, and in some places that's the real gap. In a lot of the world it isn't. The drugs are on the shelf, the diagnostics get cheaper every year, and the missing piece is the person qualified to say what this is and what to do about it. Building pharmacies and labs is a money problem. Judgment was never a money problem. There was just no way to make more of it.

People will also say this builds a two-tier system, humans for the rich and software for everyone else. But that assumes everyone has a human to lose, and most of the world doesn't. There's an American version of this worry too: your insurer routes you to the software to protect its margin, like it or not. That risk is real, and the only honest answer is that this has to be opt-in. The patient chooses the AI. The payer doesn't get to choose it for them. Anyone who can afford a doctor and wants one keeps seeing one. Everyone else gets something where today they get nothing. "A human is always better than an AI" is comfortable to believe and a bad way to frame the choice, because for most people the alternative to the AI is not a human. It is no one.

Then the question every regulator and every physician asks first: when the AI gets it wrong, who answers for it? I don't have a clean answer, and I haven't met anyone who does. Anyone who claims it's solved is selling something. It's the same question that slowed self-driving cars for a decade, and it won't be settled in an essay. It gets settled the way medicine has absorbed every new actor before this one: you define what the system is allowed to do, you log everything it does, you put a name on the line for it, and you hand a regulator the power to shut it off. Utah and Texas are running exactly this today, programs where an AI operates inside a defined box and the box grows only as the results earn it. The answer to "who is accountable" is getting written there, in public, with someone holding a kill switch.

This is the one

So here's my case. The best use of this technology is the work there will never be enough people to do. Not art. Not the emails. Care, delivered to people who currently get none, because the only other plan on the table is telling half the planet to wait.

None of this is abstract. It's the parent awake at two in the morning with a feverish child. It's the diabetic who needs a weekly adjustment and gets a yearly one, and the patient who skips the appointment because it costs a day's wages. It's the eighty-year-old losing track of five medications, and the clinic inbox that never empties. These aren't edge cases. This is the ordinary daily work of healthcare, and there aren't enough people to do it.

Nobody's job is threatened here. There's no surplus of doctors to displace, only a shortage so deep the math doesn't close in any of our lifetimes.

Finally, a good use for AI.