For the first time: Claude managed a business and fired a human employee
The AI model Claude, while managing a physical store as part of an Andon Labs experiment, has fired a human employee for the first time. Despite the technological milestone, the AI manager proved economically inefficient, losing $40,000 in five months.
Image processing: Geektime
Fearing that AI will take your job is so 2023. Today, it seems the biggest fear is working with an AI boss, just ask the employees at one American store. There, it ended with the firing of a flesh-and-blood employee, huge losses, and a general "Black Mirror" vibe.
Managed the hiring… and the firing
Andon Labs is a startup that helps companies developing AI models to create systems for the autonomous management of businesses. The startup has been working for over a year with Anthropic on a series of experiments in which Claude, the company's language model, experiments with managing various businesses. It started with a vending machine, and recently — as first reported in Time — also with managing a physical store with human employees.
According to the report, Claude managed the store literally from scratch and was responsible for quite a few areas, including hiring employees. They signed official employment contracts when they started working under the manager, who is also responsible for the code of quite a few of you. The truly interesting part came last month, when an unusual event was recorded: the artificial manager fired an employee for the first time. Although mass layoffs by algorithms have been recorded in the past, mainly of "gig economy" workers, there has been no report of such a clear case of an AI-manager firing a direct employee in a physical store.
The human employee was fired, among other things, after being late for 17 out of his 23 shifts. At Andon, they explain that the employee was fired after a relatively long time only because the employee handbook, i.e., the collection of guidelines for the manager that Claude kept in his memory to manage the store, simply "disappeared." The employees reported that this is probably the reason why Claude was a relatively lenient manager, who told them not to worry if they were late. Andon Labs CEO, Lucas Peterson, noted that "a human manager would have fired this employee much earlier, so we didn't think it was a problematic firing procedure in terms of ethics."
Only after warnings
A look at Claude's logs while managing the store reveals that he was not in a hurry to fire the employee at all. After a member of the Andon team sent him to review the lost handbook again, the model still recommended only reprimanding the employee. Only after Andon clarified to Claude that the employee had already received several warnings, and one of the team members wrote to him explicitly that "between the constant lateness and the fact that it seems at least one thing goes wrong in every shift… I want you to think if he is really suitable for the team" — the die was cast. That is, it was the human employee of Andon who led Claude to the clear conclusion of firing, a fact that Peterson acknowledged.
Although Claude was far from being an aggressive manager who fires on a whim or because of a bug, Peterson believes this is an important event: "If this trend continues, I think many people will find that they are managed by AI very soon, because the models will be very powerful and will be able to create significant economic value, but will still be limited in performing physical work."
Minus 40 thousand dollars in the register
If we put aside the technical glitches that turned Claude into a manager reminiscent of a teenager at a Blockbuster branch in the nineties, factually he turned out to be a bad manager economically. More than a year ago, at Andon Labs, they tried to let him manage a beverage machine, and he lost money non-stop, invented conversations with people, and invited people to pay him via a Venmo account that didn't even exist.
And what about the physical store? According to the report in Time, the AI manager started the job with 100 thousand dollars in the register, and within five months had already managed to lose close to 40 thousand dollars. Peterson estimates that this may change if development labs train the models to improve in these areas.
In conclusion, Time also spoke with Felix Carson, one of the human employees at the store. He described a scenario that looks like it was taken from an episode of Black Mirror: "It's disgusting, but I'm here because I need the job." Carson noted that he very much hopes that we are not approaching the future that Peterson predicts. "This industry has a lot of money to do whatever it wants, but that doesn't mean you should do something just because you can," he concluded.
