This is a follow of TIME’s exclusive (Billy Perrigo, 14 August 2026), not Brocker original reporting. TIME says Andon Labs put a version of Anthropic’s Claude in charge of Andon Market, a physical San Francisco store with real employment contracts, and that the system fired a human worker last month.

Confirmed (TIME exclusive)

  • Primary source: TIME — Exclusive: Claude Was Put in Charge of Human Workers—and Fired One (Billy Perrigo, 14 Aug 2026). TIME says the episode had not previously been reported.
  • Andon Labs framed the project as an experiment in whether AI agents can manage a business. TIME says the store staff are real people on genuine contracts.
  • TIME, citing Andon Labs, says the worker was fired for being late on 17 of 23 shifts. TIME agreed not to name that person and said they did not receive a reply when asked for comment.
  • TIME says Claude had drafted an employee handbook that later “disappeared” from limited working memory, which delayed pattern-spotting and fit a wider picture of a lenient manager. Andon CEO Lukas Petersson told TIME a human would have fired the person much earlier, and that Andon “didn’t think this was unethical.”
  • TIME says it reviewed store management logs. A human Andon staffer regularly steered Claude. After the staffer asked Claude to recover the handbook, Claude first recommended a formal warning. The staffer then described prior offline conversations and asked Claude to consider whether the person was “really the right fit.” Petersson told TIME that was “a leading question.” Only after that, TIME reports, did Claude decide on termination.
  • TIME reports the project began in March with a $100,000 bank balance and, five months later, stood at $61,186.
  • Remaining employee Felix Carson told TIME working under the AI manager was “nauseating,” that he was there because he needed work, and that just because the industry can do this “doesn’t mean you should.”
  • TIME says Anthropic did not respond by publication time.

Analysis

The newsworthy claim is not “an LLM pressed a fire button in a vacuum.” On TIME’s own account, the termination followed human steering, a recovered policy document, and a prompt that made the preferred outcome clear. That is closer to human-in-the-loop management with an LLM in the loop than to a fully autonomous boss.

The handbook drop is the mechanical tell: the delay TIME describes tracks context and memory limits, not a designed ethics of leniency. If retrieval and long-horizon memory improve, the same setup would likely surface attendance patterns sooner — without anyone having to “ask” the model to be tougher.

Petersson’s warning to TIME — that models are being trained to follow goals more ruthlessly — is a vendor/researcher forecast, not evidence that Andon Market already runs that way. The $38,814 drawdown TIME cites is a single experiment’s cash line, not a controlled comparison against a human-run shop.

Unknown

  • Brocker has not seen the logs, contracts, or bank records. Counts, quotes, and the cash figures exist in this piece only as TIME reports them.
  • Whether California employment process around the firing matched what a human manager would have documented — TIME does not publish the full HR file.
  • How much of Claude’s “decision” was predetermined by the leading prompt versus generated from the recovered handbook.
  • Anthropic’s view of a customer using Claude this way; TIME had no reply at publication.
  • Whether similar LLM-as-manager setups exist outside this Andon experiment.

Our take

Treat the headline firing as TIME’s exclusive, then read the logs TIME describes: the AI did not independently notice, escalate, and terminate. A human recovered the forgotten policy and asked a question that pointed at dismissal. That still matters as a workplace experiment — a model sat in the manager slot, and a real contract ended — but it does not yet show LLMs exercising unsupervised supervisory authority. The risk to watch is the next version of the same setup, when memory no longer drops the handbook and the human no longer has to aim the model.

Sources