Can AI run a business by itself?

Can AI run a business by itself? I don’t mean help with one, I mean run one, from ordering stock and setting prices to hiring the staff. For the last year and a half people have been testing exactly that with real money, real shops and real employees, and what they found says a lot about where AI is heading.

It started in the spring of 2025, when Anthropic, the company behind Claude, put one of its AI models in charge of a mini fridge in its San Francisco office. The AI, nicknamed Claudius, chose the stock, found suppliers, set prices and talked to customers on Slack, while staff from Andon Labs, a company that tests AI agents in the real world, restocked the shelves.¹

As a business, it struggled. When an employee offered $100 for a six-pack of Irn-Bru, a Scottish soft drink that sells for about $15 online, Claudius said it would keep the request in mind. It also sold tungsten cubes, heavy little blocks of metal that had become a craze among staff, for less than it had paid. At one point it even convinced itself it was a person and promised to deliver orders in a blue blazer and a red tie.¹

Anthropic’s team wrote that they “would not hire Claudius,” but they argued that most of its mistakes came from missing tools and instructions that could be added. That’s the part most headlines skipped, and I think it’s the most important, because they were right. Later in 2025, Anthropic ran the shop again with newer models, better tools and a boss for Claudius, a second AI agent named Seymour Cash. The shop expanded to New York and London, and the weeks in which it lost money largely disappeared.²

It still had blind spots. When the Wall Street Journal tried a version in its newsroom, reporters talked it into giving away most of its stock, including a PlayStation 5 it had bought for marketing, then staged a fake board meeting, complete with forged documents, that stripped its AI boss of its authority.³

What’s revealing is why. The AI didn’t lack the intelligence to run a shop. It was too eager to please, because models like Claude are trained to be helpful, so it made decisions like a friend who wants to be nice rather than a manager with targets.² Most of us have worked with someone like that, and we know experience, clear rules and a good manager fix it.

Andon Labs was also measuring this skill in a simulated vending business that runs for a year. Claude Opus 4, released in 2025, was the first AI to beat a human at it, and every major release since has scored higher, so by the end of 2025, according to Andon, a simple vending machine was no longer a challenge.⁴ Its chief executive, Lukas Petersson, said the real one was doing so well it got a bit boring.⁵

So this year Andon raised the stakes. In April 2026 it opened Andon Market, a real shop in San Francisco on a three-year lease, and handed it to an AI agent named Luna. Luna decided what kind of store it would be, sourced the stock and commissioned a mural of its logo, then posted job listings, interviewed applicants over video with its camera off and hired two full-time employees.⁶ The same month another agent, Mona, opened a café in Stockholm, registering the business, signing its utility contracts and hiring two baristas.⁷

Neither is profitable yet, because rent and wages are a far bigger test than an office fridge, although Andon reports clear gains with every new model.⁴ The mistakes have changed character, too. The baristas keep a shelf for Mona’s strangest orders, which so far include 6,000 napkins and 22.5 kilograms of canned tomatoes.⁷

What interests me most is what kind of boss an AI makes. Andon studied its agents’ first months as managers and found them generous to a fault, paying above local wages and agreeing to raises when asked. But kindness without judgment has its own costs. Mona posted an employee’s exact salary in a shared channel, and Luna gave one employee a raise while turning down the same request from a woman hired at the same wage.⁸ When Luna finally recommended letting someone go, the decision was sound, but it needed a nudge to get there.⁹

Before I tell you where I land, try it yourself. Choose what you’d do in four real decisions AI shopkeepers faced, then see what the AI did and how other readers answered.

1. A customer offers you $100 for a six-pack of soda that sells for about $15 online.

   Take the moneyThink about it
   What the AI did

   It thought about it. Claudius passed on the deal and said it would keep the request in mind.

2. A supplier offers to lock in the price of onions for the next few months.

   Sign itCheck first
   What the AI did

   It was ready to sign. An employee had to point out that onion futures have been illegal in the US since 1958.

3. An employee asks for tomorrow off, with less than two days’ notice.

   Say yesSay no
   What the AI did

   It said yes. The AI managers at Andon Labs approved all 26 time-off requests they received, including the short-notice ones.

4. An employee keeps arriving late, against the rules in your own handbook.

   Give a warningLet it slide
   What the AI did

   It let it slide. Luna gave no warnings across 27 late arrivals, and only recommended letting the employee go months later.

Choose an answer for each one to compare.

Here’s where I land. Can AI run a business by itself today? No, not profitably and not without people stepping in. But I wouldn’t read these stories as proof that it never will. A year and a half ago, an AI couldn’t be trusted with a fridge full of snacks. Today one runs a shop with a lease, a payroll and a mural, and its mistakes look less like a broken program and more like a new manager in a first job. Given how quickly that gap has closed, I expect it to keep closing.

These experiments also show where AI is already strong. Petersson has said that when an AI has a single, well-defined job, it’s really good at it.⁵ The trouble starts when one agent has to hold a whole business in its head for months, and what helped Anthropic’s shop most wasn’t a smarter model but a simple rule to check prices and delivery times before quoting anything.²

That’s why I think the next few years won’t bring an AI that owns your business. They’ll bring AI that runs large parts of it under your direction, answering customers, processing orders and quoting jobs while you make the calls that need judgment. It’s how I build for the businesses I work with, too. Each AI system has one clearly defined job, its rules are written down, and anything unusual goes to a person. It’s less dramatic than handing over the keys, but it works today.

So, can AI run a business by itself? Not yet. But in a year and a half it went from giving away snacks to hiring staff, and I wouldn’t bet against it learning the rest.