Why I'm Betting Everything on AI-Native Operators.

There is, somewhere in Los Angeles, a man called Matthew Gallagher, whose name I had not — until a Thursday morning in April, when I was eating a slightly underbuttered piece of toast and reading the New York Times on my phone — heard. The Times had run a profile of him. The profile contained a sentence which I have, in the weeks since, found I am unable to dislodge from my head. The sentence was, more or less, that Mr Gallagher had taken two months and twenty thousand dollars and a dozen artificial-intelligence tools and built, from his house, a company that would in all likelihood do one-point-eight billion dollars in sales this year. He had, last year, done four hundred and one million. He had, when he wrote the cheque for the company's first month, no employees. He has, at the time of writing, exactly one. The one is his younger brother, Elliot.

I put the toast down.

I put the toast down because the article — and one is not, on the whole, supposed to put toast down for newspaper articles, this being the kind of melodrama one has been trying since university to grow out of — had, at the level beneath the prose, said something which had not previously been said in print, in quite that way, by quite that newspaper, on quite that morning. It had said: the thing the founders have all been quietly muttering about, in the small bright kitchens of San Francisco, has happened. A person on his own has done it. The person is forty-one. The person is in Los Angeles. The person grew up, as it happens, in a trailer park.

I am told the toast was not, in the end, salvageable.

What he actually did.

The company is called Medvi. It is a telehealth provider, which is a nice phrase that means, in practice, that Medvi sells compounded GLP-1 weight-loss drugs to people who do not want to pay one thousand three hundred dollars a month at a regular pharmacy and would, on the whole, prefer to pay two hundred and ninety-nine. The doctors who write the prescriptions are not employed by Medvi. They are employed by another company. The pharmacies that compound the drugs are not employed by Medvi either. They are employed, similarly, by another company. Medvi is the thin, bright, almost embarrassingly thin layer that sits between an advertisement and a credit-card form, and beneath the credit-card form a small constellation of sub-contractors who do, in fact, the heavy and tedious work that companies have always done.

Mr Gallagher used, by his own account, ChatGPT and Claude and Grok to write the code. He used Midjourney and Runway to make the ads. He used ElevenLabs and a small set of agents he had built himself to handle the customer service. He created — and this is the part of his account I keep returning to — a separate set of agents whose only job was to watch the business in real time and tell him, on a Tuesday afternoon, when something was broken. He did not have meetings. He did not have a roadmap. He did not have the slightly worried look of a man who is, four times a year, performing a board update. He had, instead, a laptop, a Stripe dashboard, and the small persistent advantage of not being divided by twelve.

In its first month the company had three hundred customers. In its second, it had a thousand more. By the end of the first year it had, as I have said, done four hundred and one million dollars. He hired his brother. They are, between them, a company.

The honest part.

I want to be honest about the thing, because the thing is not entirely flattering, and the people who have written about Medvi in the manner of a sermon — there have been a great many of them — have, I think, omitted the part of the sermon that would have made it useful.

The Food and Drug Administration sent Mr Gallagher a warning letter in February. The letter was not, by FDA standards, mild. The website, in its early form, had used what appeared to be AI-generated photographs of patients who did not exist, and before-and-after pictures which were, on closer inspection, the faces of real strangers grafted on to other strangers' bodies. The customer-service chatbot had, at one point, made up drug prices, which Mr Gallagher — to his credit, and at some cost — honoured. There are class-action suits, on grounds I do not feel qualified to summarise, working their way through the California courts. Several writers, of more careful temperaments than mine, have argued that Medvi is not, in fact, the future. Medvi, they have argued, is the past, conducted at the speed of the future, with the corners cut more rapidly.

I do not, on the whole, disagree with them.

What I disagree with is the conclusion they would like me to draw, which is: therefore, the AI-native operator is a fiction. Therefore, none of this is real. This is, I think, almost exactly wrong. The corners that have been cut at Medvi were, until last spring, simply not cuttable. The fact that they can now be cut — by a man on a laptop, in a house in Los Angeles, in two months, for twenty thousand dollars — is the entire point. The story does not become less true because the man has, in fact, cut some corners one wishes he had not cut. The story becomes, if anything, more true, because it tells you with a clarity nobody asked for what the new gradient looks like and where, on it, the water now flows.

The math is dull. The implication is not.

It is the most banal observation in the world that software is becoming cheaper to make. Anyone with a browser and a sliver of patience can now produce, in an afternoon, the kind of working prototype that, in 2018, would have required a small team and a small mortgage. This is not, in itself, interesting. People have been saying it at conferences since at least 2003, by which point the saying of it had already become a tic. What is interesting is the consequence — the second-order, slightly indecent consequence — which is that the bottleneck has moved.

It used to live in engineering. It now lives in taste, and in nerve.

The bottleneck, more precisely, lives in a narrow stretch of a person's mind: the place that decides what to point the cheap, abundant software at. Everything upstream of that decision has collapsed in price. Everything downstream of it has, too. What remains valuable, almost embarrassingly so, is the pre-conscious flick of attention by which a particular person looks at a particular problem — a thirteen-hundred-dollar drug, in Mr Gallagher's case, that everybody had silently agreed to keep selling at thirteen hundred dollars — and says, with an unjustifiable confidence, this is the one I'm going to spend a year of my life on. Big teams can no longer reliably do this. Big teams have meetings. Meetings, as a class of object, are designed to launder taste through committee until nothing remains but the soft beige of consensus.

Operators — small, slightly feral, occasionally insufferable individuals with a working sense of disgust at the prevailing arrangement — can.

I have been wrong about this before.

It is fair to say that, when I was eighteen, I believed Big Things were built by Big Companies, in the way one believes, at eighteen, that suits make a man and that conferences make a city. I went to my first technology event in a hotel ballroom in Toronto. There were lanyards. There was a buffet. A man in a polo shirt told me, quite earnestly, about his roadmap. The roadmap had four lanes. The lanes had owners. The owners had owners. I remember walking out into the cold thinking I had glimpsed the engine of progress, when in fact what I had glimpsed was a slide deck.

Two years and several humiliations later — a door-to-door sales job that taught me to recognise pity at twenty paces, a dropshipping business that taught me the moral cost of margin, a painting company in Toronto that taught me you can in fact lose money employing very large men with very small attention spans — I had revised my opinion. The engines of progress, it turned out, looked nothing like a roadmap. They looked, in retrospect, an awful lot like a man at a kitchen table in Los Angeles, eating something out of a takeaway container, talking to a chatbot at half past one in the morning, watching the dashboard tick.

The man at the kitchen table is now the bottleneck. He is also, with all the small ethical complications that come with it, the moat.

What an operator actually does.

I should be precise, because the word operator has been tossed around lately by people who plainly do not operate anything more complicated than a thermostat. By operator I mean, very specifically, a person who can hold the entire stack — from the colour of the button to the line of the contract to the wording of the FDA reply — inside one head, at one time, without dropping any of the bits. It is not a technical claim. It is a claim about working memory and about caring.

An operator notices when the loading spinner is the wrong shade of off-white. An operator notices when the second e-mail in the onboarding sequence sounds, very faintly, like an estate agent. An operator notices, three weeks before anyone else, that the metric everybody is celebrating is being inflated by a small, persistent bug in the analytics pipeline. (This last is, in my experience, the single most common species of corporate self-deception.) An operator can, when pressed, write the SQL, write the copy, ring the customer, and apologise convincingly for all three. None of this is glamorous. Most of it cannot be written into a job description without sounding deranged.

It happens, however, to be the entire job.

The Mr Gallagher I have, in the last several weeks, been trying to assemble in my head from public reporting is — at the level beneath the corner-cutting, beneath the FDA letter, beneath the slightly tabloid arc of the New York Times piece — a man who held a great many of these bits in one head at one time. He is not, I suspect, a saint. He is, I suspect, an operator. The two are not mutually exclusive, and they are not the same.

The agents are not the point. The leverage is.

People become, I think, slightly silly about the agents themselves. They want to know which model. They want to know which framework. They want a scoreboard. I cannot help them. The choice of model matters, in the way the choice of pen matters to a novelist — that is, somewhat, but not as much as you would assume from looking at the pen aisle. The point is not which agent. The point is what an operator with five competent agents can now do that an operator with five competent humans could not.

The answer is: most of it, and faster, and at three in the morning, and without a single performance review.

I have a folder on my laptop titled, with no irony at all, The Staff. In it live the prompts, the tools, the small persistent memories of half a dozen agents I have built and rebuilt over the past year. One of them writes draft cold e-mails in a voice I have spent months fine-tuning until it sounds, eerily, like a slightly more confident version of me. Another reads compliance updates and shouts when something looks dangerous. A third, on whom I rely most and trust least, edits my writing — including, I will admit, the paragraph you are reading now, although the sentence about the toast was mine and I will fight for it.

The gap is the thing.

If I am betting on anything, it is the gap. Not the technology, which will commoditise; not the model, which will be replaced by a better one before this essay is twelve months old; but the gap between the people who have understood what the technology means for their working day and the people who have not.

The gap is, at present, comically wide. I sat through a meeting last month in which a perfectly intelligent person, with a perfectly impressive title, complained that her team had been unable to produce a market-research document in under three weeks. I had produced one, by accident, on the train, while waiting for a coffee. It was not better than hers. It was certainly not worse. The point is that the cost — measured in time, in attention, in the particular currency of a Tuesday morning — had simply ceased to exist for me, while remaining stubbornly real for her.

Mr Gallagher has, in his own way, made the same observation, at much greater scale and with much greater consequence. The hims-and-hers of this world — companies with, last I checked, somewhere north of two thousand four hundred employees and a dignified five-and-a-half per cent net margin — were not, in any meaningful sense, doing something he could not do. They were doing something he could do, with twenty thousand dollars and two months, slightly worse but very much faster.

This is what the gap looks like, before anyone has decided to call it a gap. It looks like a slightly puzzled woman in a meeting room, wondering why the new graduate is suddenly outshipping the senior team. It looks, in less flattering form, like a forty-one-year-old man in Los Angeles, on his second laptop of the year, doing one-point-eight billion dollars in sales while the FDA tries, via certified mail, to get his attention.

It will, very soon, look like something else.

What I am not saying.

I am not saying you should be Matthew Gallagher. I am not, especially, saying you should sell compounded weight-loss drugs at half price to people who answered an Instagram ad on a Tuesday evening. There are, in the Medvi story, several decisions I would not make and several more I would not, on the whole, defend. The deepfaked patient photographs are not, by anyone's reading, fine. The hallucinated drug prices are, at minimum, a sign of a system stood up too quickly and reviewed too slowly. The FDA letter exists for a reason.

What I am saying is that the underlying mechanic — one human being, holding the entire business in one head, deploying agents at the parts of the business that used to require thirty colleagues — is real, and is now visible in the largest newspaper in the country, and is not going to become less real because the first widely-reported example of it had, on inspection, a number of corners one would have preferred uncut.

The lesson of Medvi is not copy Medvi. The lesson of Medvi is: the leverage is here. The judgment that ought to accompany the leverage is, in many cases, not. The next great fortunes — and I do mean fortunes, in the slightly old-fashioned sense — will go, I suspect, to the operators who manage to develop both at the same time.

So.

I am twenty-something, in San Francisco, in a flat that does not, thank God, smell of fried chicken or, for that matter, of compounded semaglutide. I have four products at varying stages of disrepair, a portrait of myself in a tuxedo that I am told is unbecomingly vain, and a deep, slightly evangelical conviction — sharpened, not blunted, by reading the Medvi piece — that the next decade will reward, disproportionately, almost rudely, the small, taste-driven, agent-augmented operator who has bothered, in addition to the leverage, to develop a working conscience.

I am betting everything on it. Not because I am certain. (Anyone certain about anything in software is, almost by definition, mistaken.) But because the alternative — to wait, to credential, to hire, to wait some more, to spend three weeks producing a market-research document somebody else produced on a train — has begun, lately, to feel like a kind of cowardice dressed up as patience.

If you happen to see Mr Gallagher in Los Angeles, you might tell him, on my behalf, that I read the article. Tell him also, if you would, that the toast survived. Most of the toast, anyway. Enough of it for breakfast.

— B.