The One-Person Company With an AI Workforce

A company used to be a crowd: founders, managers, specialists, assistants, each with a desk and a salary. A new kind of company is now running with a crowd of one. The founder sits at the top of an org chart where every other box is software: agents that write the code, answer the customers, place the ads, watch the numbers, and wake the human only when something truly needs a human. This is the honest tour of that company form in October 2026: who predicted it, who has built it, its departments, its monthly staff bill, and where it breaks.

The bet: a group chat, a prophecy, and a scorecard

Start with the bet, because the bet is why everyone is watching. Sam Altman, the chief executive of OpenAI, first laid out the idea to Reddit co-founder Alexis Ohanian in late 2023, and the clip spread widely in early 2024. Altman said that he and his tech CEO friends keep a betting pool in their group chat on the first year there is a one-person billion-dollar company, something that, in his words, "would have been unimaginable without AI and now will happen." Notice what the bet requires. Not a one-person company that earns a good living; those have existed for decades. A company worth a billion dollars, run by one person, where the missing employees were replaced by AI. Three separate bars, all clearing at once.

In May 2025, the bet got a second prophet with a deadline. At Anthropic's Code with Claude conference, chief product officer Mike Krieger asked CEO Dario Amodei whether a billion-dollar business could be created by a single person using AI. Amodei said it would certainly happen, possibly as soon as 2026, then told the press the odds: a 70 to 80 percent chance in 2026. He even named the likely categories: proprietary trading, tools for software developers, and businesses that do not need salespeople and can automate customer service. Krieger added the footnote from his own life: "I built a billion-dollar company with 13 people." That was Instagram, and the main reason it needed those 13, he said, was content moderation, a job AI would now handle better.

So here is the scorecard, two and a half years into the bet. The revenue bar has effectively been hit, by a company in a category nobody named, and the next section goes inside it. The valuation bar has not been proven: the closest company has never raised outside money and has no official valuation. And the headcount bar was missed by the smallest possible margin: the closest founder hired exactly one employee, his brother. The bet is close enough that Altman, when the story broke, emailed the reporter to say it appeared he had won, and that he "would like to meet the guy." The rest of this post is about that guy, the much larger club behind him, and the fine print.

The closest anyone has come: Medvi

Matthew Gallagher, 41, works from his house in Los Angeles. In September 2024 he launched Medvi, a telehealth company selling GLP-1 weight-loss drugs, the class of medicines behind Ozempic and Wegovy. He spent $20,000 in total on the software and the first month of marketing, and it took him two months to get the company off the ground. He had no employees. What he had was more than a dozen AI tools, used the way another founder would use departments.

ChatGPT, Claude, and Grok wrote the code that powers the platform and the copy on the website. Custom AI agents, bots he built to perform tasks on their own, connected his software systems to each other. Midjourney and Runway generated the images and videos for his ads. ElevenLabs voice tools talked to customers. Other AI systems analyzed performance in real time. The parts no AI could legally do, he rented: CareValidate and OpenLoop Health, two telehealth infrastructure platforms, supplied the licensed doctors, the prescriptions, the pharmacies, the shipping, and the compliance. Gallagher kept the customer relationship: the brand, the website, the ads, the checkout, and the service. In his own summary: "It's not an A.I. company, but I did it with A.I."

The numbers, which The New York Times was given access to Medvi's financials to verify, are the reason this post exists. Medvi got 300 customers in its first month and 1,000 more in its second. In 2025, its first full year, the company generated $401 million in sales, served 250,000 customers, and kept a net profit margin of 16.2 percent, roughly $65 million of profit. This year it is on track for $1.8 billion in sales, and Gallagher told the Times the company was bringing in more than $3 million a day. His total profit so far: $70 million to $80 million. For scale, Hims & Hers, the public company that pioneered this exact business, reported $2.4 billion in revenue last year with 2,442 employees and a net margin of 5.5 percent. Medvi is running at nearly three times the margin with a headcount roughly 1,200 times smaller.

And then the fine print, because Medvi is also the best evidence of where the one-person form bends. Gallagher did hire one person: his younger brother Elliot, whose job is intercepting and filtering communication so Matthew can focus. He added two engineers on contract. In September he began assigning human account managers, seven of them, all on contract, to a subset of customers, because some customers want a person who remembers their children's names; each manages several hundred clients, using AI to carry the load. The clinical work was never his at all; it belongs to two partner companies with their own staffs. The Times delivered the formal verdict itself: Medvi is technically not a one-person billion-dollar company, since Gallagher hired his brother and has some contractors, and the company, having raised no outside funding, has no official valuation. Both things are true at once. The prophecy came true in spirit, at a scale nobody can argue with, and the letter of it is still waiting for a cleaner case.

One more detail explains why Gallagher pulled this off. His previous company, Watch Gang, a wristwatch subscription business built in 2016, had 60 employees and never turned a profit. "It just increased my costs, and then it delayed my decision-making because I had more people to deal with," he said. Medvi was designed by a man who had run the normal version of a company and hated what headcount did to his speed. The one-person company, in the founding case, is not a gimmick. It is a correction.

The million-dollar club: what pure solo actually tops out at

Medvi is the outlier that bends every average, so set it aside and look at the club underneath: solo founders running software businesses whose numbers are public because they publish their own revenue. ONEPC, a research project tracking one-person companies, updated its verified list in May 2026, including only founders whose revenue is publicly disclosed on a citable source. Every figure below is self-reported, and the list flags who secretly has help.

At the top sits Pieter Levels, the Dutch developer behind Nomad List, Remote OK, and Photo AI, at roughly $3 million a year across his portfolio, with zero employees. The portfolio replaces a single big product: Photo AI hit $138,000 in monthly recurring revenue in November 2025, Remote OK runs at roughly $2 million a year, Nomad List at roughly $700,000. His purest unit, Interior AI, brings in about $43,000 a month on roughly $300 a month of hosting, a gross margin above 99 percent. In March 2025 his browser game fly.pieter.com went from zero to a $1 million annual run rate in 17 days. Danny Postma, who launched the AI headshot generator HeadshotPro alone, reports $300,000 a month, a $3.6 million annual run rate. Marc Lou made $1,032,000 in 2025 across more than 16 small products, anchored by ShipFast, a starter kit for other builders. Tony Dinh's TypingMind brings in roughly $130,000 to $160,000 a month, and Jon Yongfook ran Bannerbear solo for years to $991,000 of revenue in 2024.

Two honest observations. First, the ceiling is real: pure solo software, run by the best operators alive, tops out around $3 million to $3.6 million a year. Nobody in the verified club is at $10 million alone. Second, look at the shape: small digital products with near-zero delivery cost, sold through an audience the founder built first. ONEPC's pattern summary is blunt: distribution came before product, and none of the million-dollar founders launched cold. The AI workforce amplifies a founder who already knows how to reach customers. It does not replace knowing.

Now put Medvi back on the chart and the gap teaches the real lesson. Gallagher is not doing Levels, but bigger. He is doing something structurally different: selling a physical-world outcome, medicine delivered to a door, into a market where customers were already spending heavily, on top of regulated infrastructure he rents by the transaction. The leverage did not come from a better agent. It came from pointing the agent stack at an existing river of money and declining to own any part of the riverbed. That is why the honest distance to a billion, measured in the final section, is shorter than the club's ceiling suggests and longer than the headlines claim.

The org chart: which departments the agents take first

Take a normal company's org chart and ask, box by box, what the one-person version does with it. The cases above give a consistent order of surrender.

Engineering goes first. It is the department agents absorbed earliest and most completely. Medvi's platform code was written by ChatGPT, Claude, and Grok under one man's direction. Levels and Postma build the same way, founder as architect and reviewer, models as the typing pool. The human keeps the decisions about what to build and whether it works; the agents do the volume.

Design and marketing go second. Medvi's ad images and videos came from Midjourney and Runway; its website copy came from the same models that wrote its code. In the solo club, marketing is the founder's audience plus agent production: the founder is the brand, the agents the studio.

Customer support goes third, with a warning label. A chatbot is the first hire of every one-person company because support is where volume arrives uninvited. It is also where the founding case's most instructive failures happened, so it gets its own discussion below. The mature pattern, visible at Medvi by September, is a ladder: agents handle the routine, contractors handle the relationships that need a memory and a name, and the founder handles whatever is left.

Operations and analytics go fourth, quietly. Gallagher built AI systems to watch his business performance in real time, the work of an ops team and a data analyst in a company of thousands. Foundra, a research firm that studies these operators, describes the working method of the successful ones: every tool is treated like a hire, with a job description, a clear handoff to the next tool, and a metric attached. The org chart is real; it is just written in software instead of payroll.

Sales barely exists. Amodei's category test predicted this: the model works where you do not need salespeople. Every company in this post sells through a checkout page or an audience, not through a sales team making calls. The one-person company does not automate the sales department. It chooses markets where there is none to automate.

Bookkeeping and legal are the last to delegate, and they graduate back to humans. Gallagher started with AI accounting tools and the online legal site LegalZoom. As the money grew, he moved to an accounting firm and a law firm. The pattern is exact: agents handle these departments at small scale, and then the amounts at stake buy back human professionals, not as employees but as firms on retainer. The department does not disappear. It changes vendors.

And the one job that got a human hire first will surprise you. Elliot Gallagher's job at Medvi is not engineering, sales, or support. It is filtering his brother's communications: deciding what reaches the founder and what does not. The scarcest resource in a one-person company is the one person's attention, and the first human hire in the founding case is a chief of staff whose entire function is protecting it. That may be the most transferable fact in this entire post.

The stack and the monthly bill

What does the workforce cost? Far less than the payroll it replaces, and the numbers are refreshingly checkable. Foundra's research puts a complete solopreneur tech stack at $3,000 to $12,000 a year in 2026, and operating margins for one-person businesses in the 60 to 80 percent range. Medvi's entire launch, software plus the first month of marketing, cost $20,000, in a market where its main competitor had raised $197 million. Levels runs Interior AI on about $300 a month of hosting against $43,000 a month of revenue.

The biggest line item is the brain itself. Anthropic's pricing page lists Claude Pro at $20 a month billed monthly, and Claude Max, the tier for people who work with the model all day, from $100 a month, with a top tier at $200. Paid plans include Claude Code, the coding agent, and Max bundles $100 to $200 a month in API credits. In plain terms, the senior engineer, copywriter, analyst, and first-line support agent of a one-person company rent, at the highest personal tier, for about the cost of a team lunch. Around that core, the typical stack adds a second model, a workflow tool to connect the agents, an email and payments layer, and hosting. Even a generous build lands in the low hundreds a month.

But read the bill honestly, because the subscription is not the whole cost. Agent costs scale with use: a support agent answering ten customers costs almost nothing, and one answering ten thousand costs ten thousand times almost nothing, which is no longer nothing. Platform risk is a cost too: every department is rented from a vendor who can raise prices, change limits, or deprecate the model your workflows were tuned on, with no IT department to absorb the migration. And the largest cost never appears on an invoice: the founder's supervision hours, which the next section counts. The stack plus the human backstop is the real price, and it is still absurdly low by any historical standard, a reduction of 95 percent or more against the equivalent staff on Foundra's figures, backed up by the verified margins in the solo club.

A day of supervision: what the human actually does

If the agents do the work, what is left for the person? The founding case answers with uncomfortable clarity. Gallagher told the Times he works on Medvi basically anytime he is not showering, sleeping, or spending time with his two children. He made an AI clone of his own voice to schedule personal appointments, to buy back working time. When Medvi's support chatbot could not handle a customer, it transferred the call to the one human backstop it had: Gallagher's personal cellphone. He fielded more than 1,000 customer service calls himself. That is the true job description of the one-person founder. You are not the boss of the machine. You are its escalation path, its exception handler, and its on-call engineer, at every hour, at every scale.

The day divides into four kinds of human work. Direction: deciding what the company does next, which product, which market, which price. No case in this post delegates this: the founder's judgment is the company. Review: reading agent output before it ships. Gallagher tested his support bot by asking it for lasagna recipes and tweaking it until it stopped offering them, a funny story that is also a precise picture of founder quality assurance. Exceptions: everything the system was not designed for. In March 2025 Gallagher changed something small on the website, went hiking, and got a call asking why there had been no orders for an hour. His update had broken ordering. With nobody else to fix it, he sprinted home; the outage cost him roughly 200 customers. Relationships: the partners, the platform owners, the customers who need a person. "At this point, I kind of want to hire people because I'm lonely," Gallagher said. The Times' headline writer earned their pay: super efficient, and a little bit lonely.

The operators who last, per Foundra's fieldwork, run a strict cadence against this sprawl: deep work batched into one or two long blocks, routine operations handed to agents and templates, and two protected blocks a week for talking to actual customers, whose words get turned directly into the next product change. The cadence matters because the failure mode of the job is not laziness. It is becoming a full-time manager of software, spending the day approving, correcting, and firefighting agent output: a middle manager in a company of one. The founders in this post escaped that trap the same way, by hiring, renting, or automating the filtering itself, starting, in Gallagher's case, with a brother whose job is deciding what deserves the founder's eyes.

Where it breaks: hallucinations, regulators, and the 3 AM problem

The failure modes of the one-person company are not hypothetical, and the founding case supplies most of them. Medvi's customer service chatbot made up prices for drugs; Gallagher honored the invented prices. It hallucinated, telling customers Medvi sold hair-loss drugs it did not sell. His early site carried before-and-after weight-loss photos taken from around the web with the faces changed, and ads the Times flatly called A.I. slop. He later replaced the photos with real customers. This is the compounding problem in its mildest form: an agent error at a normal company gets caught by the next human in the chain. At a one-person company, the error ships at machine speed until the one human notices, and the notice can come from a customer, a journalist, or a regulator.

The regulator came. In February 2026 the Food and Drug Administration sent Medvi a warning letter for misbranding: its site language falsely implied the company compounded the drugs itself and suggested FDA approval that compounded products do not have. Medvi was not singled out; more than 30 telehealth companies received similar letters in the same enforcement wave. But consider who receives that letter at a one-person company. There is no compliance department and no one whose job was to read the rules before the website shipped. Gallagher had already graduated from LegalZoom to a law firm, the model working as intended, but the letter illustrates the legal fact the whole form runs into: an agent cannot be liable. Regulators, courts, and tax authorities deal with persons, and in a one-person company there is exactly one person to deal with, for everything, including outputs of software he did not personally write or approve. His clinical partner OpenLoop Health also disclosed a data breach in January 2026 affecting at least 68,000 patients, a reminder that rented departments bring their own risks home.

Step back and the industry data agrees with the anecdotes. Gartner, in its October 2026 Hype Cycle analysis, notes that only 17 percent of organizations have deployed AI agents so far, and expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027, killed by escalating costs, unclear business value, and inadequate risk controls. Those are companies with compliance teams and managers whose only job is oversight, and two in five of their agent projects still die on governance. The one-person company has the same governance problem with a governance department of one, who is also the entire executive team. Add platform risk, a single vendor's outage or account problem can idle every department at once, and the 3 AM problem, named for the hour the site breaks and the org chart's only engineer is asleep. The form does not fail because agents are weak. It fails, when it fails, at the supervision layer, the one layer that cannot be delegated.

The honest distance to a billion, and what it means for the rest of us

So measure the distance plainly. Altman's bet had three terms: one person, a billion dollars, because of AI. By October 2026 the score reads as follows. Revenue: a two-person company is on track for $1.8 billion of sales this year, after a verified $401 million first year, and its main line runs at over $3 million a day. Valuation: unproven, because Medvi has never raised money and carries no official valuation, so the strictest reading of the bet, a company worth a billion, remains open. Headcount: missed by one brother, plus a penumbra of contractors that the honest accounting must include, two engineers, seven account managers, media agencies, an accounting firm, a law firm, and two clinical platforms. Amodei's 70 to 80 percent call for 2026 looks, at worst, a year early and a category off: he guessed trading, developer tools, and automated service, and the answer was a weight-loss drug middleman, which obeys his deeper rule almost perfectly. It is an area where the product is a transaction, the institutions can be rented, and the customer's need was already there. "It's not an A.I. company, but I did it with A.I." may end up as the epitaph of the whole prediction: the first near-billion one-person company did not sell AI. It used AI to avoid becoming a company in the old sense at all.

The leverage math, finally, in one place. Revenue per employee, computed from the figures in this post: Medvi produced about $200 million of 2025 revenue per full-time employee. Hims & Hers produced just under $1 million per employee. Midjourney, the previous lean legend, reached $200 million in annual recurring revenue with roughly 11 employees, about $18 million each. Levels produces about $3 million per employee, the employee being himself. The curve is not smooth. It says the one-person form is worth roughly $3 million a year in pure software, roughly $18 million a head in a hit consumer product with a tiny team, and a sum that breaks the scale when one person rents an entire regulated industry and points an agent staff at its customers. The distance to a billion was never about agent quality. It was about which layer of the economy the founder chose to own, and the winner chose the thinnest layer with the fattest river running under it.

For ordinary careers, which is most of us, the useful news is not the billion. It is that the same org chart now fits in a laptop at normal scale. Solo-founded startups rose from 23.7 percent of new companies in 2019 to 36.3 percent by mid 2025, and 38 percent of seven-figure businesses are now led by solopreneurs who replaced hires with AI workflows, per Foundra's compilation. The stack costs $250 to $1,000 a month at the tiers in this post, and the surviving skill is the one the founding case hired for: judgment about what deserves attention. Jobs inside companies are reorganizing around the same fact. The person who can direct agents, review their work, catch their fabrications, and stand accountable for the result is doing the founder's job description whether or not they own the company. Learn to be the backstop on purpose, and the one-person company becomes a description of your own working life, at whatever scale you choose.

Why this belongs at Starvard

The one-person company is a curriculum disguised as a headline: it asks one person to direct machines, judge markets, and answer for the result, exactly the trio Starvard teaches. How AI Works gives you the working model of the staff itself, what agents are, what they can hold, and where they fabricate. Entrepreneurship 101 teaches the founder's side honestly, testing an idea cheaply and reading the real survival odds before you bet a career on an org chart of one. And Python 101 is the quiet superpower in every case above: the founders who direct agents best are the ones who can read, and occasionally fix, what the agents produce. All three are free, like everything at Starvard.

Watch

• Lex Fridman, "Pieter Levels: Programming, Viral AI Startups, and Digital Nomad Life | Lex Fridman Podcast": youtube.com/watch?v=oFtjKbXKqbg. Three and a half hours inside the solo club's founding mind, including how Levels automates his operations and why he refuses to hire, straight from the operator himself.
• Machine Money TV, "One Baker, Zero Staff, a Team of AI Agents | Machine Money TV Ep. 04": youtube.com/watch?v=4Cx1sKWrIZA. A plain-language walk through a one-person business job by job, following the money when there is no payroll, with the Medvi and Levels evidence and its own caveats printed in the description.

Sources

• The New York Times, Erin Griffith, "How A.I. Helped One Man (and His Brother) Build a $1.8 Billion Company", April 2, 2026. The Medvi profile: launch in two months on $20,000, the verified $401 million in 2025 sales and 16.2 percent net margin, 250,000 customers, the brother as the only employee, the chatbot fabrications, the 1,000 calls to the founder's cellphone, the hiking outage, the Altman email, and the Times' own verdict that Medvi is technically not a one-person billion-dollar company.
• PYMNTS, "The One-Person Billion-Dollar Company Is Here", April 3, 2026. The Altman betting-pool quote, the full Medvi tool list including ElevenLabs, the outsourcing split with CareValidate and OpenLoop Health, and the line this post borrows for its job description: the founder becomes the sole human backstop for every system failure, at any hour and at any scale.
• Monday Momentum, Justin Wright, "The One-Person Billion-Dollar Company Just Happened", April 6, 2026. The Medvi versus Hims & Hers comparison table (revenue, margins, headcount, startup capital), the February 2026 FDA warning letter and the wider telehealth crackdown, the OpenLoop Health breach affecting at least 68,000 patients, and Midjourney's $200 million of recurring revenue with roughly 11 employees.
• Inc., Ben Sherry, "Dario Amodei Predicts the First Billion-Dollar Solopreneur by 2026", May 23, 2025. Amodei's 70 to 80 percent probability for 2026, his candidate categories (proprietary trading, developer tools, automatable service businesses), and Mike Krieger's Instagram comparison.
• ONEPC, "20 One-Person Companies Earning $1M+ Per Year (2026)", updated May 10, 2026. The solo club: Pieter Levels' portfolio figures and Interior AI's hosting economics, Danny Postma's $300,000 a month, Marc Lou's $1,032,000 year, Tony Dinh and Jon Yongfook. All figures founder-reported and publicly disclosed.
• Foundra, "What the $401M Solo Founder Story Tells Us About AI Leverage in 2026", May 2, 2026. Solo-founded startups rising from 23.7 percent of new companies in 2019 to 36.3 percent by mid 2025, 38 percent of seven-figure businesses led by solopreneurs, the $3,000 to $12,000 annual stack cost, solo operating margins of 60 to 80 percent, and the weekly operating cadence of the founders who last.
• Gartner, "Gartner AI Hype Cycle: Why Control Now Drives AI Value", October 1, 2026. Only 17 percent of organizations have deployed AI agents, and more than 40 percent of agentic AI projects are expected to be cancelled by the end of 2027 over cost, unclear value, and inadequate risk controls.
• Anthropic, "Plans & pricing", claude.com/pricing, opened October 2026. Claude Pro at $20 a month billed monthly, Claude Max from $100 a month with a $200 top tier, Claude Code included in paid plans, and $100 to $200 a month in bundled API credits on Max.

Tagged: Technology · Business · AI Agents · AI Entrepreneurship · Startups · One-Person Business