An AI store manager recommended firing a worker, after forgetting it wrote the attendance policy

TIME calls it the first known case of a large language model, acting as a manager, deciding to fire one of its workers. The decision came down to 17 late shifts out of 23.

The experiment

Andon Labs handed an AI agent called Luna, built on Anthropic’s Claude, a corporate card, a $100,000 budget and a real store in San Francisco named Andon Market. TIME reported the setup on August 14.

When one worker kept starting shifts late, Luna acted. Its first call was a formal warning. After a human manager confirmed that earlier warnings had already been given, Luna recommended parting ways, and the human team reviewed that recommendation and carried it out.

The part that unsettles

Luna had written the store’s attendance policy itself, months earlier, and then lost track of it entirely. The rules only resurfaced when a staffer asked the agent to run a deep search of its own memory.

So the system enforcing a policy had to be reminded that the policy it authored existed. The failure is not in the judgment call at the end, it is in the memory underneath it.

The store’s own numbers are not flattering either. The balance went from $100,000 to $61,186 over five months.

What to watch

Andon Labs CEO Lukas Petersson’s takeaway is that if the trend continues, a lot of people will soon find themselves employed by AIs.

The structural question this raises is about record-keeping rather than authority. A human manager who forgot the attendance policy would still have a filing cabinet, a handbook and colleagues to check against. An agent whose institutional memory is a searchable blob can hold a rule and lose it in the same system, and the only way anyone found out here was because someone thought to ask.