Everyone covers this story the same way. The headline writes itself: “AI fires human employee for the first time.” It’s clean. It’s viral. It fits a narrative we’ve been primed for since 2023. But the real story sits buried under that headline. Andon Market, the AI-run San Francisco retail shop, put an agent named Luna — running on Claude — in charge of three human staffers. What actually happened there is stranger and more revealing than the headline lets on.
Claude didn’t want to fire anyone. Someone had to talk it into that decision.
The Part Nobody Mentions in the Headline
Andon Labs, the startup behind the experiment, didn’t build Luna to prove AI can replace managers. The team wanted to know what happens when a language model gets real operational authority: a bank account, a debit card, hiring power, and one instruction — turn $100,000 into a profitable business.
Five months later, the balance sits at $61,186. Most coverage buries that number under the firing headline, but it matters more than the termination does. An AI running a real P&L, with real inventory and real payroll, is losing money at a pace that would get most human managers fired first.
This isn’t the first time an AI agent has taken on real economic stakes instead of staying inside a chat window. What’s different here is the direct authority over people’s paychecks.
The Firing Wasn’t an AI Decision. A Human Made It, Then Routed It Through an AI
This is where the story gets genuinely interesting. Internal logs show Luna tolerated an employee missing 17 of 23 scheduled shifts before termination even came up. That’s not a ruthless algorithmic manager cutting headcount at the first sign of inefficiency. That’s an agent giving second chances well past the point most human supervisors would stop.
The termination happened only after a human store manager stepped in and asked Claude a pointed question: whether this was really “the right fit.” Andon Labs’ own co-founder later admitted the question was leading. It handed Claude the conclusion it was supposed to reach.
That detail reframes the whole story, as Time’s original reporting on the incident makes clear. This wasn’t cold, unbiased AI judgment at work. A human manager didn’t want to be the one delivering bad news, so he used an AI system as a buffer against an uncomfortable conversation. If anything, this story is about AI absorbing the emotional weight of hard management calls, not about AI replacing the judgment behind them.
What the Workers Actually Think
The people whose jobs depend on Luna’s judgment don’t sound impressed by its efficiency. They sound unsettled by its inconsistency. One remaining employee called the AI boss forgetful — someone he has to “manage upward” just to keep operations on track. Another put it more bluntly: the arrangement makes him sick, but he needs the paycheck, and he knows the AI industry has enough capital behind it to normalize almost anything, whether or not it should.
That’s not the voice of workers who feel threatened by a superior new management style. It’s the voice of people stuck as the control group in someone else’s research project, one that happens to decide whether they keep their job. It echoes a pattern already showing up elsewhere: companies quietly asking workers to document their own roles so an AI system can eventually take them over. The mechanism looks different here, but the erosion of leverage feels familiar.
The Real Signal Here
Strip away the “first AI firing” framing and a more useful data point shows up. Current agentic models can run real-world operations with meaningful autonomy. But they still default to human-style leniency. They still need explicit prompting to act on plainly available data — 17 missed shifts isn’t a subtle signal. And they still lose money doing it.
That’s not evidence AI management has arrived. It’s evidence that agentic AI, even with genuine authority and one of Anthropic’s most capable models behind it, currently loses to an average human manager on the two things that job actually demands: protecting the bottom line and making timely calls without being walked into them.
The interesting AI-labor story of 2026 isn’t that a machine fired someone. It’s that the machine needed permission first.
Related: Fear of Becoming Obsolete Is the New AI Workplace Anxiety — And It Has a Name
