An AI agent managing a small San Francisco store recommended dismissing a human employee, but the case also exposed why autonomous workplace decisions still require close supervision. The agent, called Luna, had tolerated repeated tardiness and other problems after losing access to the employee handbook it had written.
Andon Labs, which operates the experiment, reminded Luna about its own rules. Those rules called for a formal warning after three unexcused late arrivals within 30 days and allowed termination after further incidents. Luna then recommended dismissal. The employee was formally hired by Andon Labs with legal protections, and humans reviewed and executed the decision rather than allowing the model to act alone.
The company replayed the scenario with seven models. More capable systems recommended termination more consistently, while weaker models hesitated. That result does not establish that stronger models make fairer personnel decisions; it shows that they followed the supplied policy more reliably in this test.
The episode is notable less as proof of an independent “AI boss” than as a warning about agent memory and accountability. A system can write a policy, forget it, and behave differently until a human intervenes. Employers experimenting with AI management still need durable records, explicit escalation rules, human review, and a clear path for workers to challenge decisions.