AI runs a convenience store
An experimental shop in San Francisco opened where an AI system manages many operations while two humans staff the floor, creating a live testbed for operational AI and human override design. The store’s setup raises questions about event pipelines, inventory sensors, anomaly detection and where to place human-in-the-loop controls (nbcnews.com).
A new shop at 2102 Union Street in San Francisco opened on April 10 with a manager that has no body, no keys, and no cash drawer. The manager is an artificial intelligence system named Luna, and the humans on the floor work for it. (nbcnews.com) (andonlabs.com) The store is called Andon Market, and it sits in Cow Hollow looking like a normal boutique selling books, candles, games, plants, chocolate, and sweatshirts. The unusual part is that Luna chose the merchandise, set prices, picked opening hours, and even decided on the mural on the wall. (andonlabs.com) (andon.market) Luna is not a robot walking around the aisle. Luna is software with a corporate card, a phone number, email access, internet access, and camera feeds from the store, which is enough to do a surprising amount of manager work from a distance. (andonlabs.com) The humans are there because stores still need hands. Andon Labs says Luna hired gig workers to paint walls and build shelving, then hired full-time staff to run the shop because artificial intelligence can place orders and make calls but cannot stop shoplifting or move boxes. (andonlabs.com) One employee, Felix Johnson, told NBC Bay Area that he answered an Indeed job ad and interviewed with Luna over Zoom. Andon Labs says Luna created profiles on LinkedIn, Indeed, and Craigslist within five minutes of being deployed and started screening applicants on its own. (nbcbayarea.com) (andonlabs.com) Customers do not tap a self-checkout screen to pay. NBC Bay Area reports that shoppers pick up a phone in the store, talk to Luna, and Luna charges them for the purchase, which turns checkout into a live test of whether an artificial intelligence manager can handle messy real-world requests in real time. (nbcbayarea.com) This did not come out of nowhere. In June 2025, Andon Labs and Anthropic ran a smaller experiment called Project Vend, where an Anthropic Claude Sonnet 3.7 system managed a mini store in Anthropic’s San Francisco office for about a month. (anthropic.com) (andonlabs.com) That earlier system could search the web, email for restocking help, track balances, set prices, and choose products, which sounds close to a basic shopkeeper’s job description. Anthropic said the point was to see how an artificial intelligence agent behaves over long stretches of time when money, inventory, and customer demand all keep changing. (anthropic.com) Andon Market is the bigger version of that idea. Instead of a refrigerator-sized office shop, Andon Labs signed a three-year lease for a real storefront and handed the operation to Luna, which makes the store less like a demo and more like a standing experiment in where human override has to live. (andonlabs.com) The company is blunt about the goal. On its website, Andon Labs says it is trying to prepare for organizations run autonomously by artificial intelligence and does not think humans will be able to keep up with every step an agent takes, which is why this store is really a safety test wearing the clothes of a neighborhood shop. (andonlabs.com) So the question in San Francisco is not whether Luna can recommend a candle or reorder chocolate bars. The question is which decisions stay with software, which decisions snap back to the two humans in the room, and how many mistakes a real store can absorb before “artificial intelligence manager” stops sounding futuristic and starts sounding expensive. (nbcnews.com) (andonlabs.com)
