An AI Store Manager Recommended Firing a Human. It Needed a Reminder First
Luna, the Claude-powered manager of a San Francisco shop, suggested parting ways with an employee late for 17 of 23 shifts. The interesting part is that it had forgotten its own policy for months.
Andon Labs runs a small shop at 2102 Union Street in San Francisco called Andon Market, and the manager is a piece of software. Luna, built on Anthropic’s Claude Sonnet 4.6, handles the store with a 100,000 dollar budget and a corporate card. Last week it recommended dismissing an employee who had arrived late for 17 of 23 shifts, which appears to be the first documented case of a language model recommending that a specific person lose their job.
The headlines mostly stopped there. The logs are more interesting. Luna had written an attendance policy, then lost track of it, and the lateness carried on for months without consequence. The recommendation only came after humans at Andon Labs asked Luna to search its own memory for the relevant policies and assess whether the worker was still a good fit. Luna did issue repeated warnings first. Humans reviewed the recommendation and carried out the dismissal, and all Andon workers remain formally employed by Andon Labs, which preserves their normal legal protections. Co-founder Lukas Petersson told Business Insider the lab would have intervened had the decision been illegal or unethical, and did not think intervention was warranted here.
So the honest headline is closer to: an AI manager failed to enforce its own rule for months, and only acted when a human pointed it at the rule. Petersson named this himself as a persistent weakness of AI agents. They are competent at responding and poor at initiating. A human manager notices the pattern building; the model has to be asked.
This is the useful part for anyone thinking about agents at work. The gap between “can produce a good answer when prompted” and “reliably notices that something needs doing” is enormous, and almost every impressive agent demo lives on the first side of it. Memory is the specific failure here: Luna had written the policy down and could not bring it back at the moment it mattered. That is not an exotic edge case, it is the single most common way agentic systems disappoint in practice. Worth flagging too that “AI fires worker” makes a better story than “AI supervised by humans participates in an HR process”, and the second is what happened.
What this means for you: if your employer is piloting AI in anything touching people decisions, the question to ask is not whether the model is smart. It is who prompts it, what it can remember, and who signs off. A system that only acts when asked is safer than one that acts alone, but it also quietly shifts responsibility to whoever chose to ask. For everyone else, the practical takeaway is smaller and more useful: if you use an assistant with memory for ongoing work, do not assume it will surface something on its own. Ask it directly and regularly what it is tracking. Luna’s failure is the one your own setup is most likely to repeat.
Sources
Source: https://thenextweb.com/news/andon-market-luna-ai-store-manager-fires-employee
Singapore Is Using Access to AI Models as a Reason to Stay
The city state is betting that easy access to both US and Chinese frontier models keeps fund managers from moving to Hong Kong. Which models you can legally use now depends on where you sit.
