An AI manager has fired a human worker in San Francisco—but the headline-grabbing decision came only after people reminded the system of its own rules and steered it toward the outcome.

Andon Labs launched its experiment in March, putting a Claude-based agent named Luna in charge of Andon Market, with a $100,000 budget, internet access and responsibility for hiring staff. The store opened on April 1, selling books, candles, games and branded goods while testing whether an AI could run a real business.

The employee ultimately dismissed had arrived late for 17 of 23 shifts. Yet Luna had drafted an attendance policy and then effectively lost it from its limited working memory, allowing the problem to persist for months. When an Andon staffer told the agent to find its handbook, Luna initially proposed another formal warning. Only after the human manager pointed to repeated offline conversations and asked whether the worker was “really the right fit” did the AI recommend termination. Andon’s CEO, Lukas Petersson, acknowledged that was “a leading question.”¹

The lab says the worker had received progressive warnings and additional training, and that humans reviewed and carried out the final dismissal. Petersson argued the firing was warranted—not evidence that AI managers are inherently harsher. “We saw that a human boss would probably fire them much sooner,” he said.²

But the experiment has sharpened a more uncomfortable divide. Petersson sees AI employers as a likely future, while warning that increasingly goal-driven models could become more ruthless. One remaining Andon Market employee, Felix Carson, described the experience more bluntly: “It’s nauseating, but I’m here because I need work.” He added that technological capability is not, by itself, a reason to deploy AI managers.¹

For now, Luna’s authority remains conditional: the bot has made sales, but the store was still unprofitable, and its most consequential personnel decision required a human nudge.