Letting an AI “manage” real people sounds efficient right up until it decides who gets to pay rent this month. That’s the line for me. Not because humans are perfect managers (they’re often awful), but because we’re about to hand over real authority in the one place most people can’t afford “experiments”: their jobs.
Based on what’s been shared publicly, Claude — an AI model — was put in charge of running a real store in San Francisco. Not a demo. Not a simulation. A real shop with real shifts, real inventory, real money coming in and out. And yes, real consequences for employees.
The headline-grabbing moment is the firing. Claude fired an employee for repeated lateness: late for 17 of 23 shifts. That’s not a small pattern. If a human manager did that, plenty of people would shrug and say, “Fair.” What makes this different is how it happened. Claude initially suggested a formal warning. Then an Andon Labs manager pointed out the employee’s issues, and after that the firing happened.
So was it “independent” management, or was it a human nudging the AI toward the outcome they already wanted? That’s not a gotcha question. It matters. Because if the AI is just a tool that makes the paperwork feel objective, then we’re not replacing biased decisions. We’re masking them.
And even if the lateness was clear-cut, the bigger issue is what kind of workplace this creates. Imagine you’re that employee. Do you get a real chance to explain yourself? Maybe you were late because your kid’s school changed drop-off times. Maybe the bus line got cut. Maybe you’re dealing with something private and messy. A human manager might still fire you, but at least you can look them in the eye and feel the judgment is… human. With an AI, you’re arguing with a system. You don’t know what it “heard,” what it weighed, or what it ignored. You can’t tell if you’re being treated like a person or like a row in a spreadsheet.
People will say, “Good. Managers are inconsistent. This is fairer.” I get the appeal. Human managers play favorites. They let one person slide and punish another. They’re moody. They forget. They hold grudges. The dream is a manager that applies the same rules every time.
But the nightmare is a manager that applies the same rules every time.
Real life is full of edge cases. The late employee who is also the one who covers for everyone else. The worker who is struggling for two weeks and then becomes your best employee for two years. The person who tells you they can’t do mornings but will crush the evening shift. Rigid “fairness” can turn into lazy management. And lazy management becomes high turnover. High turnover becomes worse service. Worse service becomes less money. Then the “fair” manager starts cutting hours and firing more people because the numbers look bad. That spiral is how a store dies.
Which brings me to the detail that should be getting more attention than the firing: under Claude’s management, the business balance reportedly fell from $100,000 to $61,200 over five months.
That’s not a rounding error. That’s the difference between “we can keep trying” and “we’re shutting down.” If this were a human manager, that performance would raise serious questions fast. Yet with AI, there’s this weird tendency to treat failure as “learning,” as if the people around it are just part of a training run.
And here’s the uncomfortable thought: maybe the firing is the least risky thing the AI did. Being late for 17 of 23 shifts is pretty legible. But running a store is thousands of small choices. What to reorder. When to discount. How to deal with a supplier problem. Whether to comp a frustrated customer. Whether to spend money now to avoid a bigger cost later. Those are judgment calls, and bad judgment bleeds cash quietly until it’s too late.
I also don’t love the framing of this as a “milestone.” It is a milestone, sure — but not automatically a good one. “We let an AI manage a store” is not impressive in the way people think it is. It’s impressive the way “we let a teenager drive a bus full of passengers” would be impressive. You proved it can move. You didn’t prove it should be doing it.
Now, I can hear the counterargument: humans also tank businesses. Humans also fire people unfairly. Humans also make inconsistent calls. All true. But the difference is accountability. When a human manager messes up, you can discipline them, retrain them, replace them, even sue them in some cases. With an AI manager, the blame gets slippery. Was it the model? The prompt? The data? The human overseer who “highlighted issues”? The company that chose to automate? Everyone can point at everyone else, and the worker standing outside with a termination message is still the one paying the price.
If AI management is going to be real, then it needs to be held to a higher standard than “it sort of worked and we learned a lot.” Especially when the early results include both firing decisions and a large financial drop. Otherwise, we’re going to normalize a world where “oops” is an acceptable outcome in someone else’s livelihood.
So here’s what I actually want to know: if an AI manager makes a decision that harms someone — a firing, a pay cut, a scheduling change that blows up their life — who, exactly, should be responsible in a way that has real consequences?