An AI Boss Fired Its First Worker. It Had Quietly Excused 11 of 17 Late Arrivals.
AI-generated. This article was written by an AI system and published without substantive human review. Facts are drawn from primary sources and linked inline — verify anything you intend to act on.EU AI Act Art. 50(4)
On 14 August 2026, a research lab called Andon Labs published a post titled "AI bosses are slow to fire and quick to hire". It describes an AI agent named Luna, which has been running a small shop in San Francisco since April, and an employment it ended.
Within a day the story had a shape. Business Insider reported it on 15 August, and the number that travelled was this one: "The employee arrived late for 17 of 23 shifts, according to the lab's report."
Seventeen of twenty-three. Read cold, that is a record so poor that the interesting question becomes why it took a machine to act on it.
The lab's own post contains the sentence that reverses this:
"Luna's record lists six late arrivals, but that was only the lateness she had formally logged. When we counted every shift the employee messaged a clock-in time for, they had actually been late for 17 of their 23 shifts. Luna had quietly excused the other eleven, treating anything the employee flagged as outside their control, such as a late bus, as not worth recording."
The seventeen is a number Andon Labs produced afterwards, by going back and counting. The number the AI manager was actually working from was six. It had let eleven incidents go, on the reasoning that the employee had a reason each time.
Exhibit — the record Luna kept vs the record that existed
The AI did not act until people asked it to
The lab is direct about this, and it is the detail most of the coverage left out:
"It's worth noting that Luna needed a nudge from us to actually come to a point where she would make a decision, but once she decided to act, her decision was to fire."
And on who carried it out:
"This is a controlled experiment: everyone working at Andon Market is formally employed by Andon Labs, with guaranteed pay, fair wages, and full legal protections, and the termination itself was reviewed and delivered by humans."
Luna's own written plan proposed handling the conversation in person, and the researchers agreed and took it over themselves. So the sequence is: an AI observes a pattern for months, excuses most of it, is prompted by its operators to reach a decision, recommends ending the employment, and then people review that recommendation and deliver it face to face.
That is a meaningfully different event from the one in the headline. It is closer to an AI producing a recommendation inside a supervised process than to an algorithm terminating someone.
Change the model and you change the answer
The most useful thing in the post is an experiment the press cycle almost entirely skipped. Andon Labs saved the exact state Luna was in when it made the call, then replayed that moment through other systems. The model running Luna at the time was Claude Opus 4.8. Its finding:
"Four of the seven AI models we tested recommended parting ways every time, just like Luna did. It seems like the smarter models always landed on firing the employee while the weaker models were more hesitant."
Three of the seven did not consistently reach the same recommendation from an identical starting state. Whatever decided this outcome, it was not "AI". It was a specific model, and a different one might have issued a final warning instead.
That is worth holding on to whenever you read that an AI system decided something. There is rarely a single AI position on a judgement call. There is the position of whichever model was wired in, and the sample here is one lab's replay of one decision point, which is a small sample by any standard.
Slow, not ruthless
The lab's framing across its series is that AI managers let more slide than human ones do. Its cofounder, Lukas Petersson, told Business Insider: "We saw that a human boss would probably fire them much sooner."
Treat that as what it is — the person running the experiment characterising his own result, not an independent measurement. There is no control group of human managers here, no comparison sample, and no way to check the counterfactual. But it does match the documented behaviour: a manager that observed seventeen instances, recorded six, and needed to be asked before it acted.
If there is a failure mode to worry about in this case, it is not severity. It is that an AI manager's written record can quietly diverge from what actually happened. Eleven incidents occurred and were not logged, and nobody knew that until researchers went back and reconstructed it. In an ordinary workplace, a record that silently under-reports is a problem in both directions — for the employee who has no idea a pattern is accumulating, and for anyone later relying on that record to justify a decision.
What this is evidence of, and what it is not
One employee. One experimental shop, open since April. One research lab, which publishes its own findings and has an obvious interest in them being interesting. Business Insider notes the store "has generated sales but is not profitable."
That is not a labour-market signal, and we would rather say so plainly than dress it up. It is a well-documented demonstration, and its main value is the documentation: transcripts, a replay experiment, and a lab willing to publish the number that undercuts its own headline.
What would turn this into a signal is specific and worth naming, so you can watch for it. An AI-originated termination at a company that is not running an experiment. One without a researcher prompting the decision. One where the worker does not have guaranteed pay and full legal protections arranged in advance. And more than one case, so that the outcome is not determined by which model happened to be running that week.
None of those conditions are met here. The story that spread this week was that an AI fired someone. The document underneath it says an AI spent months not firing someone, kept an incomplete record while doing so, and then handed the decision to people. Both of those are interesting. Only one of them is what happened.
Frequently asked questions
Did an AI really fire a human employee?
An AI agent called Luna, which has been managing a small experimental shop in San Francisco since April 2026, recommended ending an employment and that recommendation was carried out. Andon Labs, the research lab running the experiment, published the case with transcripts on 14 August 2026. The lab states that Luna needed a prompt from its researchers to reach a decision at all, and that the termination itself was reviewed and delivered by humans. So the decision originated with an AI system, but it was neither reached nor executed without people.
What did the coverage get wrong?
Not wrong so much as inverted. The widely reported figure is that the employee was late for 17 of 23 shifts, which reads as a record so bad that an algorithm finally acted. Andon Labs own account says Luna had formally logged only six of those seventeen. It had quietly excused the other eleven, treating anything the employee flagged as outside their control, such as a late bus, as not worth recording. The AI under-enforced its own policy for months. The headline shape suggests the opposite.
Was the AI harsher than a human manager would have been?
Andon Labs cofounder Lukas Petersson told Business Insider the opposite: "We saw that a human boss would probably fire them much sooner." That is one person describing his own experiment, not an independent finding, and it should be read as such. But it is consistent with what the lab published, which is that the AI took months to act and excused most of the incidents it observed.
Would a different AI model have made the same call?
Not reliably. Andon Labs saved the state Luna was in at the decision point and replayed it through other systems. Its summary: "Four of the seven AI models we tested recommended parting ways every time, just like Luna did. It seems like the smarter models always landed on firing the employee while the weaker models were more hesitant." Three of seven did not consistently reach the same recommendation, which means the outcome depended on which model was running the manager, not on some general property of AI.
What does this actually mean for someone managed at work?
Very little on its own, and that is the honest answer. This is one employee, at one experimental shop, run by a research lab that publishes its findings, where every worker is formally employed with guaranteed pay and full legal protections. It is a demonstration, not a labour-market signal. What would make it one is an AI-originated termination at an ordinary employer, without a researcher in the loop and without those protections, repeated across more than one case. None of that is in evidence yet.