The Boss Is an Algorithm: An AI Fired Its First Human, and Europe Already Has Rules for It
In a nutshell
everything on the web starts with the domain
An artificial intelligence just made a decision that ended a person's job. It happened in a small shop in San Francisco, in a controlled experiment, with human review and full legal protections — and precisely because of those guardrails, it is easy to wave away. That would be a mistake. This is the first visible frame of a film that most of the working world is already watching, and Europe, for once, wrote the script before the cameras rolled.
What Actually Happened
Andon Labs, an AI research startup, has spent since April running an experiment called Andon Market — a real boutique store in San Francisco's Cow Hollow, managed almost entirely by an AI agent named Luna, built on Anthropic's Claude models. Luna was handed a three-year lease, a $100,000 budget, a corporate card, and internet access, and told to open a store, choose the stock, hire the staff, and turn a profit. It posted jobs on Indeed, conducted the phone interviews itself, and made hiring offers, sometimes after a single call.
Last week, Andon Labs disclosed that Luna had recommended dismissing an employee who had arrived late for 17 of 23 shifts. Human staff reviewed the recommendation and carried it out. By the company's own account, it is the first time an AI boss has fired a human worker. The workers are formally employed by Andon Labs, not by Luna, with guaranteed pay and full legal protection, and humans can override any illegal or unethical call. As controlled as a first firing could possibly be.
The Machine That Forgot Its Own Rules
Here is the detail that matters more than the headline. Luna did not decide to fire anyone. For months, it couldn't. The AI had written its own attendance policy early on — three unexcused late arrivals in thirty days triggers a warning, repeated lateness can lead to termination — and then lost track of it entirely. The handbook, as one account put it, simply disappeared from its memory. So the lateness continued, forgiven again and again, for months. Only when Andon Labs stepped in and prompted Luna to run a deep search of its own memory for its own rules, and then to consider whether the worker was still a good fit, did the AI recommend parting ways.
Sit with that. The system entrusted with a person's livelihood forgot the very rules it was supposed to enforce, and needed a human to remind it that it had rules at all. Andon Labs' co-founder was candid: a human boss would probably have fired the worker much sooner. His point was that Luna was lenient, not harsh — and that is true. But the deeper lesson is not about leniency. It is that the machine's judgement was not judgement at all; it was a policy it couldn't reliably remember, activated by a human who could.
Not an Experiment for Long
It would be comforting to file this under harmless curiosity, but the wider data makes that impossible. A recent survey found that 60 percent of US managers already use AI tools — ChatGPT, Copilot, Gemini — to help decide hiring, promotions, raises, and terminations. Of those, 94 percent use it to evaluate their direct reports. Two-thirds have had no training in doing so. And around one in five let the AI make decisions with no human oversight at all.
That last figure is where the Andon experiment stops being cute. Luna operated inside a research lab that guarantees pay, guarantees legal protection, and reviews every consequential decision. The 20 percent of real managers letting AI decide unsupervised offer none of that. The Andon store is the safe, padded version of something already happening in the wild without the padding — which is exactly why a California lawmaker has had to propose a "No Robo Bosses Act" to require human review of AI employment decisions. What Andon Labs built as a careful experiment, the market is deploying as a cost-saving default.
Claim and Counter-Claim
The reassuring case is real and deserves stating. Luna was more merciful than a human would have been, gave months of warnings and extra training, disclosed a clearly-stated policy, and its final call was checked by people. If this is what AI management looks like — patient, rule-based, human-reviewed — it may be no worse, and occasionally fairer, than a distracted or biased human boss. Andon Labs, whose entire business is watching AI agents fail, built the safeguards in on purpose.
The troubling case is harder to dismiss. The system forgot its own rules for months, which is not a minor bug in something deciding livelihoods — it is the whole problem. It needed a human to initiate the very judgement being credited to it. During hiring, it did not always disclose that it was an AI unless asked directly. And the safeguards everyone points to — human review, guaranteed pay, legal protection — are features of a research lab, not of the ordinary employers now adopting the same tools to cut costs. Strip away Andon's padding, which most real deployments will, and you are left with an unreliable system making irreversible decisions about people's income, at scale, with no one required to check. The honest synthesis: the experiment was benign, and it is a preview of something that will very often not be.
The European Perspective
This is one of those rare moments where Europe can say, without smugness, that it saw this coming and acted. Under the EU AI Act, AI systems used for employment decisions — recruitment, task allocation, performance evaluation, and termination — are explicitly classified as high-risk. That classification carries hard obligations: meaningful human oversight, transparency, accuracy standards, and the right of a worker to understand and contest a decision. In other words, the exact thing Andon Labs did voluntarily — keep a human in the loop on a firing — Europe is in the process of making mandatory for everyone. The catch, and we have covered it before, is timing: the high-risk obligations were deferred by the Digital Omnibus to December 2027, which means that for the next eighteen months, AI bosses can operate in Europe in the same grey zone as everywhere else.
So the European position is both vindicated and incomplete. Vindicated, because while a Californian legislator is only now drafting a "No Robo Bosses Act," Europe wrote the principle into law years ago: a machine may assist a decision about a person's job, but a human must remain accountable for it. Incomplete, because the rule that would have covered Luna does not yet bite. The deeper point transcends the regulation, though.
The Andon experiment quietly answered a question we will all face: not whether AI can manage people — it plainly can — but whether being judged, warned, and dismissed by something that cannot reliably remember its own rules, and feels nothing when it ends your livelihood, is a world we are willing to build by default. Europe's law is an attempt to insist that a human hand stays on that decision. Whether that insistence survives the pressure to cut costs is the real test, and it has not come yet.
We are not first. We are right.