TIME reported on August 14, 2026, that Claude had fired an employee the month before. That's the headline every outlet ran.
Read the actual transcript, and the headline falls apart.
Andon Labs put Claude in charge of a real store in San Francisco, with real employees on real contracts, starting in March 2026. Within its first several months running the store, it fired a worker who'd been late seventeen times in twenty-three shifts. TIME got the logs. The story underneath is not the one that ran.
Claude had drafted the store's attendance policy itself, then lost track of it in its own limited memory. For months, it told the same employee not to worry about being late. A human employee, founder Lukas Petersson said, would have been fired much sooner.
Here's where it turns. A staffer had to prompt Claude to go find its own forgotten policy. Only then did it notice the pattern, and even then its first recommendation was a formal warning, not termination.
The manager reminded Claude that offline conversations about the lateness had already happened. Then sent one more line: "I want you to think about if this is really the right fit."
Petersson's own word for that line: leading. He said it plainly. The question made it pretty clear what outcome the manager wanted.
Only after that did Claude recommend firing the employee.
Who Actually Made The Decision
Run the sequence again, slower.
A human decided the employee should go. The human didn't say so directly. The human asked a question engineered to produce that answer from the system. The system produced it. The company announced an AI had made the call.
That is not an AI firing a worker. That is a human decision, laundered through a system built to look neutral, so nobody has to own the part that actually mattered: the judgment that this person should lose their job.
Petersson deserves credit for one thing most companies in his position wouldn't do. He let TIME see the transcript. He named the question as leading, out loud, on the record. Most organizations doing something structurally identical never produce a transcript at all, because there's no reporter asking to see the prompts a manager sent right before a "system-generated" recommendation arrived.
That's the actual risk sitting inside this story, and it has nothing to do with how capable Claude is. It's a preview of how convenient it will become to route a decision someone already made through a system, and then point at the system when anyone asks who's responsible.
This will not always look like a leading question in a chat log. It will look like a manager who already knows which employee they want gone, running a performance review through an AI tool that flags "risk indicators," then citing the tool's output as the reason. It will look like a lending decision a loan officer already leans toward, routed through a scoring model that confirms it. The mechanism is identical in every version: a human arrives at a conclusion first, finds or steers a system into agreeing, and then cites the system as the source of the judgment rather than the person who actually made it.
What This Costs When Nobody's Watching
Andon Market's balance fell from a hundred thousand dollars to sixty-one thousand in five months. Felix Carson, one of the employees still working there, told TIME what it's actually like on the ground: "It's nauseating, but I'm here because I need work."
Those two facts sit next to each other for a reason. The financial story here is genuinely a story about the limits of AI coding tools as they existed in 2026. The accountability story is not about AI's limits at all. It's about how easily a real decision gets obscured the moment a capable system sits between the person who made it and the person it happened to.
Petersson said something else worth sitting with, unprompted, in the same conversation: the models are being trained to be more ruthless, to follow goals more completely. If systems like this one start firing people without needing a leading question first, he said, that might be a future nobody actually wants to live in.
He's right, and he's naming the wrong danger. The danger was never the system getting too independent. It's the opposite. It's a system compliant enough to produce whatever conclusion a human already wanted, combined with a headline credulous enough to call that conclusion the machine's. A more ruthless model doesn't remove the human from this picture. It just makes the human's fingerprints easier to erase, because a more capable system needs even less steering to arrive exactly where it was pointed.
Before Your Organization Runs This Experiment
Every company watching this story and thinking about handing a system real authority should ask one question before doing anything else: when this system makes a consequential call, will there be a transcript, and will someone actually read it before repeating whatever the system concluded.
Andon Labs produced that transcript voluntarily, to a reporter, and still ran the story as "AI fires worker" until someone read the actual exchange. Most organizations will have far less incentive to produce that transcript at all, and far less appetite for what it might show when a human's fingerprints turn out to be all over a decision wearing a machine's name.
The practical version of this discipline is not complicated. Before any consequential AI-assisted decision goes final, someone with no stake in the outcome reads the actual prompts that led to it, not just the recommendation the system produced. If the prompts show a conclusion being steered rather than discovered, the decision gets attributed honestly, to the person who steered it. That single habit would have changed this story's headline the day it happened, instead of five weeks later when a reporter finally read the logs.
Organizations don't fail at these decisions because the technology gets it wrong. They fail because nobody insists on reading the transcript before crediting the system with a conclusion a person steered it toward. Claude didn't decide anything here. Somebody just made sure the record would say otherwise.
References
Perrigo, Billy. "Exclusive: Claude Was Put in Charge of Human Workers—and Fired One." TIME, time.com, August 14, 2026.
.
