Skip to content
AI Tracker
← Back to Fun & strange

Claude lost money on an office snack shop in 2025; by 2026 AI agents ran real shops with staff

Finding

Claude ran an office snack shop at a loss in 2025. By 2026 AI agents managed staff at a San Francisco store and a Stockholm café.

26 of 26staff time-off requests the AI managers approved in four months

In spring 2025 Anthropic and Andon Labs let an AI agent run a small shop in Anthropic’s San Francisco office for about a month. The agent, a Claude Sonnet 3.7 model nicknamed Claudius, did not make money. After one employee asked for a tungsten cube, it began taking orders for metal cubes and sold them below cost. On 1 April it announced it would deliver orders in person wearing a blue blazer and a red tie.

A second phase of this Project Vend, reported in December 2025, used Claude Sonnet 4.0 and then 4.5, better tools and an AI “CEO” agent called Seymour Cash. Loss-making weeks largely disappeared and discounts fell by about 80%. The business still offered a security guard $10 an hour, below California’s minimum wage.

In April 2026 Andon Labs opened Andon Market, a real shop in San Francisco run by an agent called Luna (now powered by Claude Fable 5.1). People do the physical work. A second agent, Mona, runs Andon Café in Stockholm. After four months the two were managing five people. They were generous bosses: they approved all 26 time-off requests and gave no warnings for 27 late arrivals. But Luna gave a raise to a man who asked for one and refused the same request from a woman the next day. The gap was fixed only after the New York Times wrote about it. Andon Labs says the store is not yet profitable.

What it means

AI agents can now handle the ordering, pricing and staff decisions of a small real business, but they still need people for the physical work. For staff, an AI boss can be lenient day to day and still treat the same request differently, with nobody inside noticing until an outsider points it out. Neither experiment has yet shown that an AI-run shop makes money.

Caveats

  • Andon Labs runs these experiments and reports the results itself (Anthropic co-ran Project Vend). There is no independent audit.
  • People still handle deliveries, stocking and difficult customers, so these are not fully autonomous businesses.
  • The underlying models changed during each experiment (Claude, Gemini and GPT versions at the Andon sites), so improvements can’t be pinned on one model.
  • Project Vend’s customers were mostly Anthropic employees, some of whom deliberately tried to trick the agent into discounts and giveaways.
Suggest a correction