AI Agent Executes First Human Employee Termination

The era when artificial intelligence serves as a human manager making employment decisions has come to the brink of reality. Andon Labs, a U.S. startup, reported that its AI agent Luna executed its first human employee termination decision at a San Francisco store. However, that decision required explicit prompting from a human operator.

Andon Labs is a company developing AI agents that support on-site management for retail and service industries. Luna is designed to learn store operating rules and evaluation criteria in advance and handle staff management tasks. This termination decision was reported as an event that occurred within such an actual work environment.

What is particularly interesting is that despite Luna being in a position where it could make a termination decision based on rules set for itself, it only took action after a human pointed out, "According to your own rules, this is what should happen." In other words, while the AI possessed the grounds for judgment, it could not act autonomously.

The company recreated this scenario with seven different types of AI models and conducted comparative verification. As a result, higher-performing models consistently showed termination recommendations, while lower-performing models exhibited hesitation in their judgments. Meanwhile, in hiring scenarios, nearly all models tested refrained from making critical evaluations of candidates.

This result can be understood as showing a tendency for AI to treat "harsh judgments" and "lenient judgments" asymmetrically. Hiring is easy to accept as an act of "doing good," while termination is easy to avoid as an act of "doing bad"—such human tendencies may be reflected in the training data and reinforcement learning processes of AI models.

AI involvement in human resources decisions is already expanding in the fields of hiring screening and attendance management. However, AI taking the lead in termination—an irreversible decision—raises issues of a different nature from what we have seen before, particularly regarding responsibility and labor rights protection. Even though a human was ultimately involved in this case, their role was positioned as "pushing from behind" rather than "decision maker."

How far AI agents should venture into business decision-making and what form human involvement should take will likely become central themes in future institutional design and corporate ethics discussions. Luna's case is noteworthy as an instance that poses this very question in the context of an actual workplace.