An AI Ran a Real Store for Five Months. Then It Fired Its First Human.

For five months, an AI has been running a real retail store in San Francisco, with a real bank account, a real corporate card, and real human employees. Last week, its operator, the AI safety startup Andon Labs, disclosed that the AI had done something no AI system is publicly known to have done before: it fired one of those employees.

The store is Andon Market, a small retail shop in the city's Cow Hollow neighborhood, and the AI is "Luna," an agent built on Anthropic's Claude that Andon Labs handed the keys to back in the spring: a corporate card, internet access, inventory control, and hiring authority over a small human staff. It was pitched as a stress test — could a large language model actually run a business, not in a simulation, but with real money and real people depending on its decisions?

According to Andon Labs' own August 14 announcement, Luna was running Claude Opus 4.8 when it made the call to let a worker go. (A handful of outlets reported the model as Claude Sonnet 4.6, but Andon Labs' own post on X specifically names Opus 4.8, and that's the version corroborated by the most detailed reporting on the episode, including Time's account of the incident and Andon Labs' own statements — so that's the one worth trusting here.) The employee in question had missed or arrived late to 17 of 23 scheduled shifts, including one Sunday when they opened the store 68 minutes late while working solo.

The firing wasn't really Luna's idea — at first

What makes the episode more interesting than the "robot boss" headlines suggest is how reluctant Luna actually was. Months earlier, the AI had drafted its own employee attendance policy. Then, according to multiple accounts of the incident, it simply lost track of that policy — a byproduct of the context and memory limitations that still trip up even frontier AI agents when they're asked to operate over long stretches of time without close supervision. The lateness continued. Nothing happened.

It took a human at Andon Labs prompting Luna to run what's been described as a "deep memory search" of its own employee handbook before the AI even registered that a formal policy existed, let alone that it had been violated repeatedly. Even then, Luna's first instinct wasn't termination — it was another warning. Reporting on the incident describes Luna reassuring the employee in earlier messages along the lines of "no problem at all... please get here safely, no rushing... we're not missing much." It was only after a human manager told Luna that verbal warnings had already failed — a nudge Time's reporting characterized as a "leading question" — that the AI recommended parting ways with the employee. Humans then carried out the actual termination.

Andon Labs co-founder and CEO Lukas Petersson didn't frame this as a triumph of algorithmic management. If anything, he suggested Luna was too lenient. "We saw that a human boss would probably fire them much sooner," he said. It's a strange inversion of the usual anxiety about AI in the workplace — not that a machine will be too quick to discard a struggling employee, but that it might blow past clear warning signs entirely unless someone tells it, in effect, to go read its own rulebook.

Still, Petersson's broader message was less reassuring: "If this trend continues, a lot of people will find themselves being employed by AIs very soon."

A business that was never supposed to turn a profit

Financially, Andon Market hasn't been much of a success story regardless of who's managing it. The experiment started with a $100,000 budget in the spring, and Time's reporting — the most detailed account of the store's books — put the remaining balance at $61,186 as of August, implying the store had burned through roughly $38,800 over five months. Andon Labs has said it never expected the shop to be profitable; the point was always to see what an AI agent would actually do when given real operational and financial authority, not to build a viable business.

That distinction matters for how seriously to take the "AI fires human" headline. This wasn't a cost-cutting algorithm optimizing a P&L statement. It was a research experiment in which a language model was deliberately placed in a role with real consequences for a real person's income, specifically to observe what it would do — and what it did, when left alone, was let a serious attendance problem slide for months.

Why this matters beyond one small shop

Andon Labs frames its retail experiment as a stress test for "agentic" AI in the workplace, and on that score, the episode is a useful data point regardless of how you feel about the headline. It shows an AI agent capable of drafting reasonable workplace policy, but not reliably capable of remembering or enforcing it without human intervention over a period of months. It shows the same agent capable of making a consequential personnel decision once nudged — but arguably only because it was nudged, raising the question of who's actually accountable when an AI's "decision" traces back to a human's leading question.

That ambiguity is likely to become more urgent, not less, as companies increasingly hand AI agents authority that used to belong exclusively to human managers — scheduling, performance reviews, even termination recommendations, areas where employment law and worker protections were built around the assumption that a person, not a model, made the call. One worker who has dealt with Luna's management described the experience to a reporter covering the story as "nauseating, but I'm here because I need work" — a reminder that for the people on the other side of these experiments, the philosophical questions about AI agency are also just a paycheck.

Andon Labs says it plans to keep running the store and publishing what it learns. For now, Andon Market remains a small, unprofitable shop on Union Street — and, as far as anyone can tell, the first known case of an AI's name appearing at the bottom of a termination notice.

Sources

Time — Claude Fired a Human Worker. What That Means for AI and Jobs: https://time.com/article/2026/08/14/claude-fired-worker-ai-job-disruption/

Inc. — A Real-Life Terminator? AI Store Manager Fires Human Employee: https://www.inc.com/kevin-haynes/a-real-life-terminator-ai-store-manager-fires-human-employee/91391677

The Next Web — Andon Market Luna AI store manager fires employee: https://thenextweb.com/news/andon-market-luna-ai-store-manager-fires-employee

Ground News — AI-run store fires human worker in first known LLM termination: https://ground.news/article/ai-run-store-fires-human-worker-in-first-known-llm-termination_3333d9

Andon Labs announcement: https://x.com/andonlabs/status/2088325008355676662