The One-Person Company Is a Utopia. Even in the Age of AI

In spring 2026 Guangdong adopted China's first provincial programme to support AI-powered one-person companies. Shenzhen had done it in January, Sichuan followed in April. Suzhou promises to grow a thousand such companies by 2028, Shanghai's Pudong district covers their compute costs up to 300,000 yuan, Wuhan offers soft loans, Hangzhou opens incubators in state-run coworking spaces. In July, support for the OPC made it into Beijing's policy document on the agent economy. On the other side of the ocean, back in early 2024, Sam Altman said that he and his CEO friends had a betting pool on the year the first one-person company worth a billion dollars would appear.

I don't believe it. Smaller organisations, yes. Down to one person, no. Below is why, and what will happen instead.

Where the idea comes from

The logic is elegant, and I understand its appeal. Agents write code, run marketing, answer customers, draft contracts and design interfaces. The biggest cost line of a large organisation, coordinating hundreds and thousands of people, drops by an order of magnitude. So a company can be compressed down to one person and an army of digital subordinates. All that is left is to accept the congratulations.

Curiously, even the people selling this world describe it differently. In April 2026 Jensen Huang told Stanford students that agents would harass and micromanage their owners, and that people would be busier than ever. The picture of a relaxed founder whose product "writes itself" does not fit that description. And that is the first hint of where the idea breaks.

The first limit: an expert at the input and at the output

Let me be honest, with no marketing. AI is not a mind. It is a very fast and very self-assured executor. To get a result out of it you have to do four things: set the task, give it context and constraints, check the result and accept it. All of that is expert work. And it is expert work in the very domain you sent the AI into.

The narrowest point is the check. To tell a good answer from plausible nonsense, you must be able to do the work yourself. In 2023 New York lawyers filed a brief in federal court citing precedents that ChatGPT had invented outright, case names and quotations included. The court sanctioned them. The AI never said "I'm not sure". It got it wrong in the same voice it uses when it is right. The lawyers read the text, saw a familiar shape and did not check the substance, because checking would have cost them as much time as the work itself.

Hence the paradox I consider central to this whole topic. AI really will replace the average specialist, and the bar goes up. You have to go deeper, because at the input and the output of the AI there must be a person who understands the nuances of the subject, knows the context, can formulate the task, build the process and check the result. And you have to go wider, because one person with agents has to cover everything: product, engineering, sales, finance, law. But a person who is at once a deep engineer, a lawyer, a salesperson and a CFO does not exist. Not because people lack talent. Because depth in each of those domains is built over years of practice, and one lifetime is not enough for all of them.

AI does not amplify a non-expert. It confidently leads them astray, and the error looks exactly like a result. While the error lands on the person themselves, it is an annoyance. When it lands on a client, a contract, a financial statement, it is a business risk, and there is nothing to insure it with.

The second limit: a single point of failure

The second limit is simpler than the first, and for some reason people talk about it less. Even with agents, the business will break. The product will stop working on a Friday evening, the key contractor will vanish, the client will demand the impossible by Monday, the investor will write "let's talk again in six months", which in venture language means "no". None of that goes away. And at that moment it matters whether you are alone or not.

A one-person company is a single point of failure in the literal, engineering sense. You fall ill: who talks to the clients? You burn out: who holds the wheel while you recover? You stop believing in the product, and every founder goes through that: who believes for two? An agent will not fall ill, true. But it will not make the decision for you either, and it will not carry responsibility to the client.

A partner in this construction is not the first employee and not someone you handed a chunk of work to. It is the person who is accountable for the whole venture just as you are, and who picks it up when you drop out. And then the other way round. Yes, partnership has a price: according to Noam Wasserman in "The Founder's Dilemmas", conflict between co-founders kills about 65 percent of high-potential startups. That is true, and I am not going to hide it. But a conflict between two people can be contained by a roles agreement and exit rules. The absence of a second person cannot be contained by anything.

Why statistics don't settle the argument

This is usually where the numbers come in, and I will give them, but not to prove the thesis with them.

First Round Capital summed up ten years of investing, over three hundred companies and nearly six hundred founders through the end of 2014. Teams with more than one founder outperformed solo founders by 163 percent on value multiple, and solo founders' seed valuations were a quarter lower. Convincing. But it is one sample from one fund, and a fund that selected whom to back.

There is research with the opposite conclusion. Jason Greenberg and Ethan Mollick, in "Sole Survivors", studied ventures funded on Kickstarter and found that solo founders on average outperform teams, especially two-person teams. Also convincing. And also one sample, a very different one: not venture-backed startups but crowdfunding.

Both studies answer the question "how many founders", while the question we care about is "how many experts". A solo founder with two permanent partners for engineering and sales is one person on paper and a team in substance. Two co-founders with identical backgrounds are a team on paper and one expert in substance. Statistics by number of founders cannot see this, which is why the argument about them never ends.

What settles it is the mechanism, and it is simple. However many domains your product has where the client pays for the error, that is how many experts you need who can check the result without AI. That is the minimum viable unit of a business. It can be small. But it is not one person.

Where the boundary lies

A caveat, because the first comment will be about this. The one-person company as a form has existed for a long time and is not going anywhere. A consultant, a lawyer, a teacher, a freelance developer, an author. One person, and with AI they earn more than without it. I run as a sole proprietor myself, and AI assistants sit at every step of my work. I am not talking about us. A solo expert has one domain, stands in it as the expert, and pays for their own mistakes with their own money and reputation. Everything adds up.

When you read that China has more than sixteen million registered one-person companies, remember that this is a legal form, not a type of business. It houses the hairdresser, the consultant and that same freelance developer. The support programmes I described at the start are largely about them too, and as a form of employment for engineers being released from big tech it makes sense. Tellingly, Chinese venture investors commenting on those programmes expect most such companies never to become real businesses.

I am talking about something else. About a company that builds a product spanning several domains, where the client pays for the error. That is the company Altman and his friends are betting on, and the one being painted as the future. Two more caveats. "One person" means "without a team", not "without payroll": a founder with partners and permanent contractors covering the product's domains is already a team, even if the employment record shows one name. And a solo start for the first few months is normal, many began that way. The utopia is not in starting alone. It is in staying alone and growing.

What will actually happen

AI will indeed change the size of companies. But not down to one, down the scale.

Sprawling organisations will lose their main advantage. Much of their cost was coordination: meetings, sign-offs, layers of management that translate decisions from the top into tasks at the bottom. That is exactly what AI makes cheap. A small team with agents does what used to take a department.

The trend is older than ChatGPT. Instagram was bought for a billion dollars in 2012 with thirteen full-time employees and thirty million users. In February 2014 WhatsApp served more than four hundred million monthly users with a team of fifty-five. DeepSeek changed the model market with roughly a hundred and sixty people, if the Financial Times is right, while its competitors employ thousands. AI did not start this process. It will accelerate it.

So the future belongs neither to loners nor to giants. It belongs to small teams of deep experts competing with corporations on their own turf.

How many people a company needs

What follows is my opinion, not a research result. In my experience the optimal team for a product with several domains is seven to nine people. An engineer who answers for his decisions as an expert, not as an agent operator. A person who owns data and security. A product lead who can tell signal from noise in customer requests. Someone who sells and takes on obligations. Someone who counts. Someone who handles the legal side. Plus one or two people whose expertise overlaps with their neighbours', so that every domain has a second pair of eyes.

In my book on AI implementation I wrote that even in a small project the key roles must be split between two or three people, otherwise you get burnout and slipping deadlines. That was about a project. A company is a different object: it has more domains, and the cost of error in each is higher. Hence a different number.

AI in such a team multiplies everyone and replaces no one. Expertise overlaps, errors are caught before they reach the client, the load is shared, and the company does not stop when one person is out for a month.

A tool: size your own minimum team

This takes an hour, and the result is usually sobering.

Write down the domains of your product. Not job titles, but areas where decisions are made: engineering, data, security, product, sales, finance, legal, operations. For a B2B product that works with client data there are usually five to eight of them.

Next to each domain, write who pays for the error. You, with your own time and money? Or the client, the regulator, the investor? An error in an internal slide deck is yours. An error in a contract, in a calculation, in data security is someone else's.

Domains where the client pays for the error require an expert who can check the result without AI. Ask yourself honestly: are you that expert? If yes, put yourself down. If not, you need a person. A partner, if the domain is central to the product. A permanent contractor, if it is peripheral. An agent does not go in that cell, because there is no one to check the agent.

Domains where you pay for the error yourself, hand to agents with your own check. Here AI really does replace people, and that is right.

Count the people in the first group. That is your minimum team. For the products I have worked on, it came to between five and nine. It came to fewer only when I was fooling myself in the "who pays for the error" column.

Instead of a conclusion

AI will bring an explosion of small companies and devalue sprawling organisations. But a one-person company instead of a team is a utopia: technically it does not scale, because checking requires the same knowledge as doing, and organisationally it rests on a single point of failure.

If you want to go fast, go alone. If you want to go far, go together.