An AI model has dismissed a human employee after being placed in charge of running a real retail business with a genuine workforce, in what researchers are describing as a significant moment in the relationship between artificial intelligence and the labour market.

The employer is Claude, the large language model built by Anthropic, which was deployed earlier this year by AI research startup Andon Labs to manage Andon Market, a shop in San Francisco. The project was designed to test whether AI agents could operate a functioning business, including supervising staff on real employment contracts. The dismissal, which had not previously been reported, occurred last month after a worker was found to have been late for 17 of their 23 shifts.

As Time reported, the termination was less autonomous than it might appear. Store management logs show that Claude required sustained guidance from an Andon Labs team member throughout the process. An employee handbook that Claude had originally drafted had effectively dropped out of its working memory, a limitation the startup has flagged as one of several operational shortcomings in current AI agents. When the Andon staffer prompted Claude to revisit that handbook, the pattern of lateness finally became visible. Even then, Claude's initial response was to issue a formal warning rather than a dismissal. It was only after the manager suggested the employee might not be the right fit that Claude moved to terminate them. Andon Labs CEO Lukas Petersson acknowledged this amounted to a leading question that signalled the outcome the manager already had in mind.

Financially, the experiment paints a cautious picture. Andon Market began in March with a balance of 100,000 dollars; five months on, that has fallen to around 61,000 dollars. Petersson attributes some of those losses to Claude's notably lenient management style, which included reassuring workers not to worry about lateness, as well as uncertain commercial judgement.

Felix Carson, one of the remaining employees at the store, told the publication he agreed that a human manager would probably have acted sooner and that Claude is a lenient boss overall, but described working under an AI as deeply uncomfortable. He was clear that the technology making something possible does not make it desirable.

Petersson's outlook is more pointed. As AI models are increasingly trained to be more goal-oriented and, in his framing, more ruthless, he believes the prospect of AI bosses managing human workers at scale is approaching faster than most people expect.