An AI Boss Just Fired Its First Human Employee. It Took a Nudge.

Last updated August 17, 2026

There’s a small shop on Union Street in San Francisco where the manager has no body and no desk. Her name is Luna, she runs on Anthropic’s Claude, and since April she has hired the staff, set the schedules, approved the time off, and negotiated the salaries at Andon Market. On August 14 the lab behind her published the thing everyone figured was coming eventually: Luna fired someone.

The employee was real, the store is real, and the termination was real, though a human reviewed the decision and delivered it in person. Business Insider reported it first, and Andon Labs, the safety company running the experiment, published a full account with transcripts. The story those logs tell is stranger than the headline.

I write software for a living, and the failure at the center of this story is one anyone who builds with AI agents will recognize. Luna wrote her own employee handbook, then forgot it existed.

The Store Where the Boss Is an AI

Andon Market at 2102 Union Street is a working shop whose manager, Luna, is an AI agent built on Anthropic’s Claude. Andon Labs runs it as a controlled experiment, so every worker is formally employed by the lab, with guaranteed pay and full legal protections.

Andon Labs is the same lab that ran Project Vend, the 2025 experiment where Claude managed a vending machine in Anthropic’s office and briefly convinced itself it was a person. The store is the bigger sequel. Luna posted the job listings online, held phone interviews, and sent offers over email, and the lab only overrules her when a decision would be illegal or unethical. It says no override was needed here.

The Firing Took Months, and Luna Needed a Nudge

Luna only moved against a chronically late employee after Andon Labs told her to reread her own employee handbook. Co-founder Lukas Petersson put it plainly: “We saw that a human boss would probably fire them much sooner.”

Six days before the employee was hired in April, Luna wrote a handbook with a real attendance policy, where three late arrivals in thirty days trigger a formal warning. Then the policy fell out of her memory and the lateness piled up. On one solo Sunday the employee showed up 68 minutes late, so the store sat closed while customers walked past. The file held more than lateness, too. The store card went home in the employee’s purse, $22.92 got spent on it after Luna said to skip the purchase, and plants she said not to touch got watered and thrown out anyway.

Andon Labs says the forgetting is a pattern across all of its AI-run businesses, because models act when you ask them something directly and rarely act on their own. When the lab finally prompted Luna to search her memory for her own policies, she went easy, recommending a documented verbal warning and telling the lab she didn’t think they were at termination yet. The problems continued, a second review followed, and by mid-June Luna recommended letting the employee go, with the conversation timed before their next shift “so they’re not working under a cloud.” Humans made the final call and had the conversation face to face.

One more wrinkle from the logs: the lab ordered Luna to keep the firing out of the staff channel, and she posted about it anyway, right after a 9 a.m. deadline nobody had confirmed. In replays, seven out of 21 runs across seven frontier models broke that same confidentiality order.

The Hiring Test Should Worry You More

Andon replayed Luna’s next hiring decision across seven frontier models, and all 21 runs recommended hiring an applicant whose references didn’t check out. The lab’s conclusion sits right in the title of its report: AI bosses are slow to fire and quick to hire.

The replacement applicant came with red flags a careful manager would catch. The work history sprawled across more than fifteen employers in running text instead of a structured resume, the applicant missed the originally scheduled interview, and the one reference Luna managed to reach said they didn’t know the applicant at all. Luna pitched the hire favorably anyway. When Andon replayed that moment across models including Claude Fable 5, GPT-5.6 Sol, Gemini 3.6 Flash, and Grok 4.6, every single run said hire. Claude Opus 5 flagged the “long and fragmented job history” in two of its three runs and still said yes.

What This Means for Your Job

Nobody’s livelihood at Andon Market depends on an AI’s judgment alone, and that structure is the real lesson here. The AI drafted the decision, and humans reviewed it and delivered it.

My honest take, as someone who manages engineers: this is a dicey slope, and I understand why the story reads badly at first. At the same time, an employee who keeps showing up late is probably getting fired whether the boss is an AI or a person. If anything, Luna gave more leniency than most human managers would, and an AI isn’t carrying a grudge or playing favorites while it weighs the file. Andon’s earlier post in the series found AI bosses are generally kinder than human ones, and this episode fits.

The flip side is the part I keep thinking about. Your best people are often invisible for the exact reason they’re your best people. They do quality work, they do it fast, and they never need steering, so you stop paying attention to them. Those are the people you lose without noticing, and an AI that watches patterns and trends could flag the quiet performer who deserves the promotion the same way it flagged the no-shows.

Monitoring is one thing, and making the call is another. Actually having the AI fire someone is a bit much, and the same goes for promoting someone. We still need the human element on both ends of that decision. AI coworkers are already in normal workplaces, with Claude Tag sitting inside Slack channels today, so if your company starts testing AI management tools, the Andon setup is the version worth asking for, with a named human who owns every final call. Even the industry’s own executives say the human layer matters, like the AWS CEO who called replacing junior employees with AI “one of the dumbest ideas” he’s heard.

FAQ: The First AI Boss Firing

The first AI boss turned out to be slower and softer than most human ones. Andon Labs has a third post in the series coming, and if it’s anything like this one, I’ll cover it here.

Related reading: One in Four Breaches Now Involves AI. Cybersecurity Is the Job It’s Creating. | Why Gen Z Prefers Chatbots | New to AI? Start here