What happened

On 14 August 2026, Andon Labs published an account of its AI-run shop experiment in San Francisco. Its agent, Luna, had recommended dismissing an employee after the researchers prompted it to revisit earlier problems.

The company made an important distinction: Andon Labs employed the staff, and humans reviewed and delivered the termination. The report concerned an earlier incident, not a dismissal that necessarily occurred on publication day.

The researchers also described gaps in Luna’s handling of policies and its assessment of a potential replacement. This was a single controlled business experiment reported by its operators, not proof that AI can reliably manage employment decisions.

Why it matters

A recommendation about someone’s work can influence a consequential decision even when software does not execute it. A confident summary may shape which facts a manager notices, which explanations they request and which options they consider realistic.

An employer considering AI support for management should begin with the evidence. Does the record distinguish an observed event from an interpretation? Can the employee explain missing context? Is the reviewer seeing the underlying material, or only an AI-generated account of it?

Take a hypothetical attendance dispute. A system might identify repeated late arrivals. A manager would still need to establish whether records are accurate, whether expectations were communicated and whether relevant circumstances have been considered. The summary cannot answer those questions merely by sounding complete.

A meaningful review allows the person responsible to reject the recommendation and investigate further. Approval should not become a ceremonial click.

The bigger shift

The boundary between administrative support and managerial authority deserves explicit design. Summarising meeting notes, proposing a rota and recommending dismissal carry different consequences. They should not inherit the same approval process because they appear in one application.

I would ask HR and operational leaders to specify which tasks AI may support and which decisions remain with named people. Include rules for correcting records, controlling access to employee information and retaining the reasons behind a decision. Have appropriate specialists assess the rules for the organisation’s circumstances.

Employees also need to know how to reach a human when a system misunderstands them. That route should be understandable and usable without having to argue with another automated tool.

The CEO’s guide to AI governance connects these decisions with broader organisational accountability. The key question is who has the authority, competence and time to exercise it.

My take

The phrase “AI boss” is memorable, but it can obscure the people making decisions about the experiment itself: what information the agent receives, what it is asked to optimise and when someone intervenes.

I would begin management applications with low-consequence support and inspect the quality of the underlying records. Before expanding, test cases involving conflicting accounts, incomplete information and a person challenging the system’s conclusion.

Judge the process on whether it produces fair, explainable and correctable decisions. Faster paperwork is a secondary benefit.

An organisation cannot delegate its relationship with an employee to a convincing interface. If AI contributes to a management decision, a responsible person must still understand the evidence, own the judgement and be available to the person affected.