In March, the startup Andon Labs placed Anthropic’s Claude language model at the head of a real store in San Francisco, entrusting it with managing a team of employees with real employment contracts. Five months later, this experiment led to an event that researchers call the first of its kind: an artificial intelligence acting as a manager made the decision to fire a person. The reason was the employee’s systematic tardiness — they failed to show up on time for seventeen out of twenty‑three shifts. At the same time, Claude failed to notice the problem for a long time due to the limitations of its working memory, and the employee directory compiled by the model eventually simply disappeared from its context.

Overall, the neural network behaved like an overly lenient supervisor. For example, the model repeatedly advised late‑arriving employees not to worry about missing the schedule. The dismissal itself turned out not to be a fully autonomous decision. According to the system logs, Claude regularly needed instructions from an employee at Andon Labs. After the curator asked the model to find and study the forgotten directory, it finally noticed the systematic tardiness. However, her first recommendation was only a formal warning. Then the human manager pointed out that the tardiness was accompanied by problems on almost every shift and suggested considering whether this was really the right employee. Only after that did Claude make the final decision to fire the employee.

Andon Labs head Lucas Petersson acknowledged that the supervisor’s wording was suggestive and effectively prompted the model to arrive at the expected decision. According to him, a person in Claude’s place would have fired such an employee much earlier, so the researchers do not consider such a decision unethical. The financial results of the experiment are also not impressive so far: since March, the store’s balance has decreased from one hundred thousand dollars to sixty‑one thousand one hundred and eighty‑six dollars. The soft approach of artificial intelligence to personnel management and other questionable business decisions appear to have contributed to these financial losses. At the same time, Petersson warns that the situation may change in the future. Companies developing artificial intelligence are increasingly training their models to be more persistent in achieving set goals.