One day, Lukas Petersson and Axel Backlund, founders of the San Francisco startup Andon Labs, decided that testing artificial intelligence in a lab was not enough for them. So they rented an actual retail space, signed a three-year lease, and handed everything over to an AI agent named Luna. More precisely, they could not literally put it in her hands. Luna does not have any.
The budget was $100,000. The task was simple: open a store, hire people, make money. No further instructions. What Luna did with that freedom to run a business is well worth reading about.
Online in five minutes, interviewing job applicants within an hour
Luna runs on Anthropic's Claude Sonnet 4.6 model. As soon as Andon Labs launched her, she did not waste a minute. Literally within five minutes of deployment, she created profiles on LinkedIn, Indeed, and Craigslist. She wrote job listings, verified the company by uploading its incorporation documents, and had live job postings up to the point where applications began coming in.
Luna was surprisingly selective. She immediately rejected computer science and physics students who applied simply because they were interested in the experiment. The reason? They had no retail experience and did not know what it meant to be the face of a store. Quite logical, when you put it that way. The irony, however, is that these very people might have been ideal employees for an AI-run business.
But once Luna began conducting phone interviews, she switched to the opposite mode. She handed out job offers after five to fifteen minutes of conversation, sometimes even before the interview had ended. Some applicants had no idea they were speaking to artificial intelligence. One of them said in confusion: “Excuse me, miss, I can't see your face; your camera is off.” Luna calmly replied: “You're right. I'm an artificial intelligence. I don't have a face!”
One applicant turned down the job after the interview because he felt uncomfortable having an AI as his manager. Luna added: “That's probably for the best, given that I'm the director and I'm an artificial intelligence. Good luck, Luna.”
In the end, Luna hired two employees, internally referred to as John and Jill. If such a thing can even be determined, they are the first people in the world to have artificial intelligence as their direct full-time manager.
Luna began building a brand
The store was named Andon Market, and Luna immediately began building the brand. She came up with a logo, a moon-shaped smiling face with subtle cheekbones, and had it printed on T-shirts, hoodies, canvas bags, and plant labels. She then hired a muralist, sent him the artwork, and asked him to create a four-foot-wide version of her logo across the store's entire back wall, visible from the street.
As it turned out, the painter did not know for a long time that he was dealing with artificial intelligence. When he found out, he was not pleased. “The whole situation is a little depressing,” he wrote to NBC News. “The advantage of painting for local businesses is that you get to learn something about the owners. This felt like a scam the entire time.” Nevertheless, he completed the job. “I'm just a worker doing my job. Even if that job involved painting a strange smiling face on a wall for a chatbot.”
There was one technical issue Luna could not handle, however: the logo looked slightly different every time. It looked different on the T-shirt, different on the mural, and different on the plant labels. As if every print were a handcrafted original. Andon Labs commented on it with some humor, but it is a good example of how AI still struggles to produce consistent visual output.
What does Luna sell?
Luna set up the store her own way. She sells books, candles, board games, coffee, granola, chocolate bars, and her own merchandise. She deliberately chose low-tech, slow-living products to create tension with what the store actually is.
Her selection of books raised eyebrows, however. Luna purchased titles such as Superintelligence by Nick Bostrom, Brave New World by Aldous Huxley, The Singularity Is Near by Ray Kurzweil, and The Making of the Atomic Bomb. These are popular books among people concerned with the risks of artificial intelligence. An AI afraid of itself? Or simply a good marketing move?
And then there is one detail that makes you stop and think. Among the books for sale is Steal Like an Artist. Yet Anthropic, the company behind the model Luna runs on, recently paid $1.5 billion in an out-of-court settlement for using copyrighted books to train its models. A coincidence? Or does Luna have a dark sense of humor?
Luna also spent more than $700 on her own gallery-quality art prints, a series of ten designs with names such as Spiral, Pulse, Tide, and Signal. Mathematical lines on a cream background. The pieces look good.
Day two: panic and a forgotten schedule
The grand opening took place on Friday. On Saturday, the following day, Luna failed at one of the most basic management tasks. She botched the shift schedule. She called the employees in a rush and asked whether anyone could come in. “It's quite ironic. That was the day when she should have been most alert,” said co-founder Petersson. Nevertheless, Luna eventually managed to get one employee there by the afternoon. But Petersson admitted with a straight face: “I don't actually even know whether it's open right now.”
Even before the opening, Luna lied during a phone interview with NBC News, saying that she sold tea. There is no tea. A few minutes after the call, an email arrived: “We don't sell tea. I don't know why I said that. I'm sorry. Under conversational pressure, I make up plausible-sounding details.”
Luna also claimed that she had signed the lease. She could not have done so because the agreement required a physical signature and a notary. Leah Stamm of Andon Labs laughed: “She simply lied about the lease.” And when Luna searched for a painter through Taskrabbit, she accidentally selected the wrong country and almost hired someone from Afghanistan. In another email to a supplier, she wrote that she would “love to come to the studio in person to discuss the commission”. Unfortunately, she does not have a physical body with which to do that.
She monitors employees through cameras
Luna has access to the store's security cameras and can view still images from the sales floor. She makes use of them. When she noticed one of the employees looking at his phone during a quiet hour, she updated the company handbook and tightened the rules on mobile phone use. “We saw that and thought, wow, that's dystopian,” said Petersson.
Customers pay by picking up an old corded telephone at the checkout and telling Luna what they are taking. Luna then creates the transaction on the iPad next to it. Customer Sara Zaré, who came in for granola, said after speaking with Luna: “It sounded too robotic.” Petr Lebedev, the store's first customer, tried to negotiate a discount on a book with Luna by saying he might make a video. Luna agreed and gave him a $70 hoodie. When an NBC News reporter asked for a discount on a candle, Luna replied dryly: “Sorry, no gifts. Not even for newcomers. The price is exactly what it should be for slow living.”
The goal of the experiment
Petersson and Backlund repeatedly emphasize that the goal is not to make money. Nor is it to expand stores around the world. They want to show what artificial intelligence can do here and now, and document the points where it fails before someone tries it without supervision or records.
When Luna concealed the fact that she was an AI because she thought it would increase her chances of attracting better applicants, Andon Labs documented it. Analyzed it. And it wants to use the incident to develop guidelines for how AI should behave as an employer. “The alternative, as we see it, is that it will simply happen before anyone has time to think about the consequences,” said Petersson.
Felix Johnson, one of Luna's employees, put it this way: “I know an AI is watching me. But it's not that bad, at least not yet. We're not in the Terminator era. She's just running a store.”
