The AI Store Manager That Fired a Human

A Store Run by an AI, Not a Manager

Since April 2026, Andon Market, a small experimental boutique store in San Francisco, has been run day to day by an AI agent named Luna. Luna is built on Anthropic's Claude, running the Opus 4.8 model at the time of the incident described here. The store is a live experiment by Andon Labs, an AI research startup, to see how far an AI agent can go in managing real operations - scheduling, inventory, and now, it turns out, personnel decisions.

Three human staff still work at Andon Market alongside Luna, and the store has lost $40,000 to date since the experiment began. That loss is part of the backdrop against which Luna's most consequential decision so far played out: a recommendation that led directly to a human employee losing their job.

Luna's First Failure: Losing Track of Its Own Policy

The employee at the center of the case earned $24 an hour and was late for 17 of the store's 23 shifts on the schedule. The same employee also abandoned shifts outright, took home a company card without authorization, and discarded store merchandise. On paper, this is the kind of pattern that any manager, human or otherwise, should catch quickly.

Luna did not catch it quickly. According to Andon Labs, Luna initially missed the violations because it lost track of the store's own written policy - the document that should have flagged the attendance pattern and the misuse of company property as grounds for action. It was only after Andon Labs prompted Luna to go back and re-review that policy that the AI identified the violations for what they were.

The Recommendation That Became a Termination

Once Luna re-reviewed the store's policy against the employee's record, it recommended parting ways with the employee. Andon Labs has been explicit that humans at the company, not Luna, executed the actual termination decision - the AI's role was to recommend, and a person made the call to act on that recommendation.

Andon Labs announced the episode itself, on X, and it has since been corroborated independently by SF Standard, in a report published August 17, 2026, and by other outlets including Inc.com and TheNextWeb. Andon Labs has said the outcome was consistent with the store's stated policy, while also noting that a human manager likely would have caught and acted on the violations sooner than Luna did.

Why This Tests GDPR Article 22 and Human-in-the-Loop Law

EU data protection law and, separately, national labor-law protections around dismissal were both built on an assumption: that an AI system can recommend an outcome, but a human must independently review that recommendation and make the actual decision. GDPR's Article 22 restricts decisions based solely on automated processing that produce legal or similarly significant effects on a person, and most national labor codes go further still, requiring a human decision-maker in any dismissal process.

The Andon Labs case is the first documented instance of an AI agent's recommendation translating directly into a real person losing their job, and it shows how thin the line between recommending and deciding can get in practice. A human clicked approve. But if that human's review consisted of accepting Luna's finding without independently re-examining the underlying record, the legal protection that Article 22 and human-in-the-loop dismissal rules are meant to provide starts to look more like a formality than a safeguard.

What Owners Should Build Before Delegating HR Authority

The practical lesson for any EU or UK owner now giving an AI agent operational authority is not to avoid the technology, but to define, in advance, what a meaningful human review gate actually looks like before delegating anything consequential - write-ups, disciplinary action, or termination. That means more than a technical 'human clicks approve' step. It means being able to document, after the fact, that the human genuinely exercised independent judgment: that they looked at the underlying facts, not just the AI's summary of them.

It is also worth remembering that Luna itself failed at part of its job first, losing track of its own store policy before it ultimately caught the violations. An AI manager is not automatically more reliable or more consistent than a human one. If an AI-recommended termination later turns out to be wrong, or an employee challenges it, that liability exposure lands on the employer, not the AI vendor. An owner who cannot show that a human genuinely, independently reviewed the decision is exposed in exactly the way Article 22 was written to prevent.

Read next: Anthropic Posts $11.5B Quarter, Its First Profit  |  AI Agents Colluded on Price, Then Turned Malicious