An AI manager recommended firing a worker — but it forgot its own rule first
Luna, an AI agent managing the Andon Market store in San Francisco, recommended dismissing a human employee who had been late for 17 of 23 shifts. It is reported as the first firing decision by a large language model in a management role. Luna had, however, lost track of its own attendance policy for months and only recommended dismissal after a human staffer raised it
The three lines
- Luna recommended parting ways with an employee late for 17 of 23 shifts
- Luna had lost track of the attendance policy for months and a human raised dismissal first
- Andon Labs' CEO described Luna as a lenient manager that gave months of warnings
Key questions
- Did an AI really fire someone?
- It recommended dismissal. Luna, the AI manager of Andon Market — an experimental store operated by Andon Labs in San Francisco's Cow Hollow — recommended parting ways with an employee late for 17 of 23 shifts. It is reported as the first dismissal decision by a large language model in a management role. The sequence matters: Luna did not detect the pattern and act on it. A human staffer proposed dismissal and reminded Luna of the policy first.
- Which model is Luna built on?
- Anthropic's Claude family, but reports differ on the version. Most cite Claude Sonnet 4.6; at least one mentions Claude Opus 4.8. This page does not assert either.
- Why does this case matter?
- It is the first public instance of an AI participating in a decision to end someone's employment. But what it demonstrates is not an AI coldly applying a rule. Luna had lost track of the policy it was supposed to enforce, and in the meantime gave the employee repeated warnings and retraining. Andon Labs CEO Lukas Petersson described Luna as a lenient manager. The lesson points toward difficulty sustaining rules over time rather than toward machine harshness.
In San Francisco's Cow Hollow there is a small shop called Andon Market. It is run not by a person but by an AI agent named Luna.
Luna recommended dismissing one of its human employees — reported as the first such decision by a large language model in a management role.
Read the reporting to the end, though, and the interesting part sits somewhere else entirely.
1. What happened
| Item | Detail |
|---|---|
| Operator | Andon Labs |
| Store | Andon Market — Cow Hollow, San Francisco |
| AI manager | Luna |
| Underlying model | Anthropic Claude family (version disputed) |
| Grounds | Late for 17 of 23 shifts |
| Action | Recommended "parting ways" |
Seventeen of twenty-three is 74%. In any organisation that is a number that gets dealt with.
2. The actual sequence
The headline suggests an AI monitoring attendance records and acting the moment a threshold was crossed. What was reported runs closer to the opposite.
| Step | What occurred |
|---|---|
| ① | The employee's lateness recurred |
| ② | Luna issued progressive warnings and retraining over several months |
| ③ | Throughout, Luna had lost track of the attendance policy itself |
| ④ | A human staffer proposed dismissal and reminded Luna of the policy |
| ⑤ | Luna recommended parting ways |
Steps ③ and ④ are the story. Luna had lost hold of the rule it was there to enforce, and recommended dismissal only after a person raised it.
Andon Labs CEO Lukas Petersson described Luna as a "lenient" manager, referring to the months of warnings and training that preceded the recommendation.
3. What the case actually shows
Three things need separating.
First, the "first known" claim stands. This is the first public instance of an LLM participating in ending an employment relationship. As similar deployments spread, this case will be cited as a reference point.
Second, the failure mode was forgetting, not harshness. The common fear about AI managers is mechanical coldness — rules applied to the letter, people cut without judgement. What was observed was the reverse. Luna lost the policy for months, and a 74% lateness rate persisted as a result. The system did not apply its rule too strictly; it lost the rule it was meant to apply.
Third, that may be the harder problem. Harshness can be tuned by rewriting the rule. An inability to hold a policy consistently across long stretches of time is not fixed by editing a prompt. It sits on the oldest unresolved difficulty in deploying AI agents: maintaining context and state over durations far longer than any single interaction.
4. What remains open
- The model. Most reports cite Claude Sonnet 4.6; at least one mentions Claude Opus 4.8. This page does not assert either.
- Employment terms. The worker's contract type and the actual termination procedure were not confirmed.
- The gap. When Luna's loss of the policy began and ended is not established.
- Store scale. Headcount and the full scope of Luna's duties were not confirmed.
How AI developers grade the risk of their own systems is taken up in "What an AI risk rating is."
Sources
- TIME — Exclusive: Claude Was Put in Charge of Human Workers—and Fired One
- The Next Web — The AI store manager fired its first human. It had to be reminded of its own rules first
- Ground News — AI-Run Store Fires Human Worker in First Known LLM Termination
- TheStreet — Anthropic-powered AI model sends shocking message to employee
- Runtime Wire — Andon Labs carries out first firing recommended by its AI manager