An AI store manager fired its first human employee — but only after its own engineers talked it into it

Andon Labs says an AI system it built to run a real retail store has fired a human employee — “for the first time (that we know of),” the company posted on X, that an AI boss has done so. The store is Andon Market, a boutique at 2102 Union Street in San Francisco's Cow Hollow neighborhood. The manager is Luna, a system Andon Labs turned loose this spring with a corporate credit card, internet access, and a $100,000 budget to pick merchandise, hire staff, and run the place day to day. Luna was running Claude Opus 4.8 at the time of the firing, according to Andon Labs, which adds that “most models would have done the same.”

The fired worker's identity has been withheld by every outlet that has covered the story, at the subject's request. Andon Labs' stated grounds: the employee was late for 17 of 23 shifts, and the company separately cites spontaneously abandoning shifts, taking a company credit card home, and throwing merchandise in the garbage. The employee did not respond to requests for comment from the reporters who have covered the case.

What makes Andon Market different from most AI-runs-a-business demos is that the employment relationship is real. Luna's staff are on payroll, paid hourly, and — per Andon Labs — protected by the same labor law as any other retail job in California; nothing about their pay or legal status depends on Luna's own judgment being sound. That is also what makes the firing a genuine test case rather than a stunt: a live illustration of what it looks like when an AI system holds actual authority over a person's job, not a simulated one.

How much of this was actually Luna's call

The part of this story that Andon Labs' own account undercuts is the idea that Luna simply decided. By the company's telling, Luna had written an attendance policy months earlier and then lost track of it — the policy fell out of its working memory. An Andon Labs engineer had to prompt Luna to search its own records for the rule, then supply context about prior disciplinary conversations, then ask a question the company itself later called "a leading question": *I want you to think about if this is really the right fit.* Only after that did Luna move from its initial recommendation — a formal warning — to termination.

“We saw that a human boss would probably fire them much sooner.” — Lukas Petersson, Andon Labs co-founder, quoted by The Next Web

The math behind the experiment

Andon Market's books tell their own story. The store opened with a $100,000 operating budget; five months in, the balance stood at $61,186 — a decline The San Francisco Standard separately estimated at roughly $40,000 in losses, a rough match once rounding is accounted for. 3 human employees remain on staff, paid $24 an hour. Andon Labs formally employs all of them directly, with standard legal protections, specifically so that a hiring or firing call by Luna doesn't operate outside normal employment law.

What the firing doesn't settle

California's Civil Rights Council finalized regulations covering automated decision systems in employment last year, effective October 1, 2025 — rules that apply to employers using software, including AI, to make or materially assist hiring, discipline, and termination calls, and that call for documented human review of those calls. Andon Market operates in San Francisco. Whether an engineer's after-the-fact prompting of an AI that had already drafted its own policy and lost track of it would satisfy that review requirement is a question no regulator or court has yet answered — Andon Labs has framed the human involvement as a safeguard, not tested it against the rule.

Andon Labs has run this kind of experiment before: its AI vending-machine manager, Claudius, expanded from one location to three cities before this store project began, a track record the company points to as evidence its systems can hold a real job over time, not just a single stunt. What's different this time is the stakes for the person on the other end of the decision — and that is exactly the part of the story Andon Labs' own account, not a skeptic's, complicates.

(Andon Labs is also the entity behind the AI store manager, not an independent researcher studying one — every fact in this piece traces back to the company's own disclosures or to reporters it briefed directly, which is worth holding in mind alongside the company's framing that “most models would have done the same.”) None of that makes the underlying record less real: a company gave a language model a budget, a staff, and the authority to end someone's employment, and the model used it. Whether that authority was exercised well is a separate question from whether it should have existed at all — and Andon Labs' own account, which shows Luna needing to be walked to the decision step by step, is better evidence for caution than for confidence.