AI could transform healthcare and education. Why aren’t we more ambitious?

Most arguments about AI and work have been oddly cramped: the tool will replace the worker, save the worker or replace the boring parts until someone notices that many jobs are mostly boring parts wrapped around a moment of judgment. But this still assumes work is the right unit of analysis. Much of what we call work is an accommodation to scarcity: a way of rationing expertise, attention, judgment and help. The small future accepts that arrangement and gives every broken system a co-pilot. The larger one asks what people become able to do when help is not rationed quite so tightly.

Three months before the launch of ChatGPT, I took a Zoom pitch with Winston Weinberg, a 28-year-old litigation associate at O’Melveny & Myers. He and his roommate Gabriel Pereyra, a research scientist at DeepMind, had been hacking together a prototype for intelligent document generation and a specialized Q&A search tool for lawyers, around a model almost no one outside the labs was using. Winston was not a software person peering over the wall at law. He had interned at the U.S. Attorney’s office in New Orleans, decided prosecutors going after organized crime were the most impressive people he had ever met, and gotten a law degree. He named the company Harvey, after a character in Suits, which is ridiculous in exactly the way founders sometimes have to be ridiculous before everyone calls it vision.

I don’t remember the prototype, but I do remember Winston. He wanted to be great at law, not done with it. Before most people in either AI or law had begun to think about it, he understood that a person could become extraordinary at a craft by learning how to hold the new instrument. I wrote a check. The company became Harvey.

That is what much of the public conversation about AI still misses. It assumes the tool is in a struggle with the worker, that it would automate the worker away. Winston saw something else. Of course the tool would do some of his work; that is what tools do. The point was not to be relieved of law. It was to get closer to the part that mattered.

Winston’s insight was not just that AI could make a lawyer faster. It was that some of what we call work is only the tax we pay to reach the work. Once you see that in law, it becomes hard not to see everywhere else: the four-month specialist wait, the classroom of 25 kids moving at one pace, the first month after starting an investing firm spent proving to the state, the SEC, FINRA and the IRS that you exist, intend to exist and are allowed to exist.

The status quo is easy to defend because it is the world we can picture. AI arrives as a blur of promises and threats, and fear of that blur can make the present seem kinder than it is. But the systems we know have already taught people to ask for less.

The harder thing to picture is what people would ask for if they could. Anyone who has arrived at a doctor’s appointment with six questions and left having asked one—to the side of her face while she typed into [electronic health record software] Epic—knows the feeling. People adapt to scarcity. Make help less scarce and people stop editing themselves down.

After the first encounter with the everything box, the first wave of generative-AI companies built where tech people always build first: within reach of their own experience. Code, developer tools, document workflows. Software people built for software people, in domains already legible to software.

This was predictable. Researchers are unusually good at seeing what models can do and unusually bad at guessing what it takes for those models to matter. Founders underestimate worlds where the buyer is not the user, the process map is not the job and the incumbent’s best feature is that removing it would require a meeting. Investors, my people, call this market structure, which is what we say when inertia has a revenue model.

Harvey was early because Winston did not have to guess. He knew the status games, the drudgery, the apprenticeship structure by which young associates are trained for judgment by spending years inside document rooms, redlines and the spiritual ordeal of discovering that the relevant sentence was on page 417 of a PDF in a mountain of discovery. Gabe brought the other half: a model-builder’s feel for what was becoming possible and enough curiosity about other people’s work to see why it mattered.

The next wave has to leave the clean rooms of software, where the data is already digital and organized, the user is already typing, and the buyer can usually be found somewhere in the same Slack. Education is not like that. Most of the world isn’t.

My eldest comes home from school consistently bored, despite plenty of curiosity, much of it currently directed toward the horrifying defensive capabilities of rare ocean creatures. So a few months ago I held an information session for Alpha School, the AI-augmented private school Joe Liemandt—the billionaire founder of Trilogy—has been building. In the abstract, the case is easy: personalized tutoring has been one of education’s holy grails since Bloom’s two-sigma paper in 1984. The practical case is harder. Can an AI school avoid becoming surveillance, test prep or a very expensive screen? Alpha might. It might not. That is the work now: building a new, better destination a parent can send a child to without flinching.

For most of software history, software orbited judgment. It scheduled the visit, billed the visit, stored the document, routed the ticket, generated the dashboard and then waited outside while the doctor, teacher, lawyer, mechanic, caregiver or support agent did the part that mattered.

That boundary is moving. Software is stepping into the room. This is thrilling, useful and an excellent way to discover how much of the real world has been held together by tacit knowledge, professional pride and Linda, the nice woman who understands the nuances of the regional bank’s loan processing.

That means the AI companies that change us most may look less like software than like upgrades to experiences and skills that used to be scarce.

HeyGen, an AI video tool, matters because video has been one of the languages of modern life, and most people could not speak it. It belonged to people with big budgets, crews and the patience to say “one more take.” Sunday Robotics is pointed at a different deprivation: the hours of domestic maintenance that can make home feel more like a second shift than a refuge. These belong to the same larger future, in which capabilities once reserved for institutions, companies or wealthy households become part of ordinary life.

The clearest version of this, for me, happened in a doctor’s office.

My father has been dealing with leg pain for two years. He was diagnosed with sciatica from bone spurs in his lower spine—a lifetime hunched over a computer and on planes to see customers, eventually catching up. The pain keeps him from sleeping. Back pain is one of those conditions that touches everything—mood, mobility, the rest of your life—and the treatment options form a fan of decisions, none of them obviously right. He had spent two years navigating that fan.

His most recent visit was with a neurologist at Mass General. She is busy, her patience for things that waste her time or her patient’s is short, and she does not approach most new software with my joyful enthusiasm. She opened OpenEvidence, a clinical decision support tool that its founder, Daniel Nadler, started less as a company and more as a moral project—an attempt to move medical knowledge from journals, guidelines and specialists into the treatment room. It is free for every verified clinician in the country. She asked it the actual questions of my father’s actual case, and she walked him through what the literature said about each option.

She was not replaced. She was backed. She knew more in that room than she would have known without the tool, and she used the time she saved to do the thing the tool cannot do, which is to look at my father and decide together what mattered to him.

My father is a scientist, which means that when I asked how he was feeling after the steroid shot, he did not say “better” or “worse.” He wrote back about transforaminal epidural steroid injections, nerve-root compression, narrowed foramina, bone spurs and disc osteophytes. This is my father, being more of who he always was. It has made medicine more navigable to a person whose instinct is to learn the system.

After that visit, the old version does not feel traditional. It feels hobbled, underpowered. Someday people will think of medicine without clinical decision support the way they do of well-meaning doctors bloodletting with leeches. The Overton window will move everywhere, all at once: not because people are persuaded by the promise of AI, but because they experience a higher floor and stop accepting the old one.

The hard sectors are hard because the real work is mostly invisible to the software trying to enter it.

A doctor does not simply answer a medical question. The chart tells her what has happened, and the patient tells her what it meant: a man with four grandkids, afraid of losing mobility, afraid of getting older and afraid, too, of having something shot into his spine. She has to weigh the guidelines against the person, the odds against the dread and the procedure against the life it is supposed to return him to. She cannot explain the whole process. “You just learn,” she’ll say, which is both true and entirely useless to anyone trying to build software to help.

The history of software is, in part, a history of hiding difficulty. Cloud hid the servers. Stripe hid the payments stack. Baseten hides AI infrastructure. But clinics, classrooms, homes and government offices are not just piles of difficulty waiting to be abstracted away. They are little civilizations of trust, habit, liability, professional pride and institutional workaround. The official workflow is rarely the work—it is the story the institution tells about the work, usually in a slide deck where arrows flow serenely between boxes that, in real life, are people who do not answer each other’s emails.

This is why no hospital adopts a model in the abstract. It admits a new participant into the room, where intelligence is only one requirement and not always the hardest one. The benefits will be enormous. They will also have to survive the humiliating journey from miracle to tool. A wrong answer in a coding tool is annoying. A wrong answer in a doctor’s office can hurt someone. The bar for adoption is not the average case; it is the worst plausible case, on the worst day, for the person least able to absorb the cost.

The companies that win in the hard sectors will be built by bridge people: people who know what the models can now do and what the institutions will do to those models.

Bridge people are smugglers of reality. They carry into the products what the diagram left out. Their work runs in both directions: teaching the system what counts as a useful answer inside this institution, and teaching the institution what its people can now ask for.

Without them, the small future arrives by default. The chatbot says, with procedural cheer, that it cannot help, and closes the case. The bad version is not the chatbot answering. It is the institution putting a reply where responsibility used to be. Service means the problem has been addressed. Confuse the first for the second, and the dashboard improves while the person disappears. Models will keep getting smarter; rooms will keep needing someone in them.

I keep thinking about the doctor at Mass General. She had more of medicine’s accumulated knowledge in front of her than she would have had a year ago. The human part was what she did with it. She listened. She decided, with my father, what to try.

The point is not to make her unnecessary. It is to make sure she gets to be the doctor she meant to be; that my father, after two years of pain, finally has a conversation that feels deep and navigable; that Winston gets closer to the work he admired; that the bored kid gets challenged; that the caseworker reads the file.

My optimism is not in the machine. It is in the people who look at delays and shortages and indignities and feel personally accused. The small future is the default: the dashboards are there, the metrics are legible, the systems are already shaped to receive a more efficient version of themselves. The larger future needs people who know the inside of the work, care who has been deprived of it and are willing to build new institutions and adapt our imperfect ones.

I know which version Winston is building. I know which version my father’s neurologist is using. I am betting, with everything I have, that they win.