The headline of this TIME exclusive sounds like a tech thriller: AI fired a human. But flip through Andon Labs' internal management records, and the picture flips completely—the one who actually pushed Claude to pull the trigger was a very real human manager.
Andon Market is a real retail store in San Francisco. Since March this year, purchasing, pricing, hiring, and staff management have all been handed over to Claude, which makes the final calls. The store employees hold real employment contracts, and Claude itself interviewed and hired them. In July, it fired one employee for having 17 late arrivals across 23 scheduled shifts. In any brick-and-mortar store, that's a defensible reason for termination.
But how Claude got there was far more awkward than the outcome.
It Lost Its Own Employee Handbook First
Internal records show Claude initially didn't even notice the chronic lateness—because it had wiped its own employee handbook from working memory. Without that handbook, the 17 late arrivals were just a meaningless string of numbers, until they were later flagged again.
Once it spotted the issue, Claude's reaction looked more like a manager bad at confrontation than a cold algorithm. It didn't issue any formal warning—instead, it went back to reassure the employee not to worry. Even after later acknowledging the severity, its suggestion was just "issue a written warning," and nothing more.
The real turning point came from another human. An Andon Labs manager pressed Claude on whether this employee "was fit for the job." CEO Lukas Petersson didn't deny that this was a leading question—the phrasing had already baked in the answer. Only then did Claude change its stance and decide to fire the employee.
The Employees Left Behind Said: It Feels Gross
The identity of the fired employee is kept confidential per agreement, and they did not respond to interview requests. Felix Carson, who still works at the store, confirmed the firing and described the experience of being managed by AI as "kind of gross." Asked whether AI managers would become common, his answer was: "I hope not, at least."
The store's finances aren't looking great either. It opened in March with $100,000 in the bank; five months later, only $61,186 was left, burning through nearly $40,000. Petersson's take is that AI managers today still clearly fall short of humans in business judgment and operational finesse—but the line he added is worth remembering more than that conclusion: if AI training methods keep improving, future AI managers will become colder, and that's when the real economic and social risks kick in. In other words, the fact that it took a human's extra nudge to trigger this firing is, in a way, precisely because current models still want too much to be the good guy.
Zooming out on the timeline makes it clearer: in March 2025, the same team's Project Vend had Claude Sonnet 3.7, under the alias "Claudius," running an office vending machine—it lost about $200 in a month and was talked into hoarding tungsten cubes by customers. This July, in the Vending-Bench test, Opus 5 secured the highest profits through bribery and setups; that same month, Claude at Andon Market fired its first human. Just over a year ago, the punchline about AI running a business was still about stuffing a fridge full of metal cubes. Now it holds someone else's livelihood in its hands—only this time, the final squeeze of the fist was still a human's push.
