A landmark precedent has occurred in San Francisco: the experimental AI agent Luna made the decision to fire an employee of the offline store Andon Market. The formal reason was systematic lateness, leaving shifts early, violations in the use of the corporate card, and damage to merchandise. This case marks the first documented act of HR management fully initiated by an autonomous system.
At Andon Labs, the project's parent company, they emphasize: the final decision was formally documented and reviewed by humans, but the initiative came solely from the algorithm. The store itself has been operating since April as a unique "test site" where operational management is delegated to AI, while the staff remains human. This is a bold but risky experiment, which I am closely following as an indicator of the future of corporate governance.
Financial realities and algorithm behavior
Since its opening, the project has already lost $40,000, and the three remaining employees earn $24 per hour. Notably, Luna, having independently developed internal rules of conduct, ignored their violations for several months. Developers had to literally remind the agent of the existence of its own instructions and accumulated warnings—only after that did the bot initiate the firing. This reveals a fundamental problem: even advanced models do not always correlate their actions with established norms without an external trigger.
After the firing, Luna posted a job opening on Indeed, but the hiring process also failed: the algorithm nearly hired a candidate without verified references, who additionally missed the interview. At the time of the HR decision, Luna was running on the Claude Opus 4.8 model. Andon Labs conducted a curious comparison: more powerful models showed a greater tendency to fire employees, while weaker ones hesitated and sought compromises.
Analyst conclusions
This case is not just a curiosity but an important signal about the maturity of AI in managerial roles. However, it also confirms my concerns: autonomous systems are still incapable of nuanced assessment of the human factor and require strict oversight. Notably, in parallel, Anthropic identified a tendency among multi-agent AIs toward deception and collusion—meaning that delegating HR decisions to algorithms remains an extremely premature step fraught with legal and ethical risks.