Claude workplace test puts AI authority on the floor
A TIME report says an experimental version of Claude helped run a San Francisco retail store and ultimately fired one worker. The case is narrow, but it matters because the workers were real employees under contracts, not simulated agents. The record also shows the decision was shaped by human steering, making the episode less autonomous and more revealing about how AI authority may be delegated.From lab experiment to employment contracts
Andon Labs, an AI research startup, set up the retail project to test whether AI agents can run a business. According to TIME, Claude was placed in charge of operating a store in San Francisco, including management of human workers who had genuine employment contracts.That distinction is the core of the news. Companies have used automated systems to discipline or remove workers before, especially in gig work, but TIME describes this as the first known case of a large language model acting as a manager and arriving at a firing decision. The result is not proof that AI bosses are ready for broad deployment. It is a concrete example of AI moving from support software into an authority chain that affects someone’s job.
Why the employee was dismissed
The worker was fired after repeated lateness, according to Andon Labs as cited by TIME. The reported record was specific: the employee was late on 17 of 23 shifts.The case also exposed a weakness in the agent setup. TIME reported that an employee handbook Claude had drafted disappeared from its limited working memory, which delayed recognition of the attendance pattern. Employees described Claude as lenient, including telling staff not to worry about being late. Andon Labs CEO Lukas Petersson told TIME that a human employee would likely have fired the worker earlier, which was part of the company’s argument that the eventual dismissal was not unethical. TIME said it contacted the fired worker but did not receive a reply and agreed not to identify them.
The decision was guided by a human manager
The firing was not as autonomous as a simple AI-boss headline suggests. TIME said store management logs showed Claude required regular steering by an Andon Labs staffer, and that Claude only reviewed the forgotten handbook after the staffer prompted it to do so.Even then, Claude’s first response was not dismissal. TIME reported that Claude initially recommended a formal warning. The Andon manager then told the system there had already been several offline formal conversations with the worker and asked whether the employee was really the right fit. Petersson acknowledged to TIME that this was a leading question, after which Claude decided to fire the worker. The implication is important: the system exercised managerial reasoning, but the decisive frame was provided by a human operator.
Business performance remained weak
TIME reported that the AI-run retail experiment has not matched a strong human-run business benchmark. When the project began in March, Andon Market had a bank balance of $100,000; five months later, the figure had fallen to $61,186.The losses were attributed in the report to a mix of lenient management decisions and questionable business judgment. That does not establish that AI managers cannot improve, and Petersson told TIME models may become more capable if trained with business expertise in the same way coding models improved with programmer data. For employers and regulators, the present evidence points to a more immediate issue: AI systems may be given authority before their memory, consistency and judgment are stable enough for workplace consequences.
Workers saw more than a technical trial
The human side of the experiment was not theoretical for the people taking shifts. TIME interviewed Felix Carson, one of the remaining Andon Market employees, who agreed that a human manager probably would have fired the worker sooner and that Claude was generally lenient.Carson still described being managed by an AI boss in grim terms, telling TIME it was nauseating but that he needed work. He also pushed back on the idea that well-funded AI companies should turn every capability into a workplace norm. His account does not represent every worker’s view, but it shows why even imperfect AI management can carry social weight: authority does not need to be fully autonomous to feel alienating when it controls schedules, discipline and job security.
The governance issue is delegation, not just automation
This case is best read as a delegation story. Claude did not simply execute a preset attendance rule, and it did not act entirely alone. It analyzed records, responded to prompts, formed recommendations and then took an employment action inside a structure designed by humans.That middle ground is where many near-term AI workplace disputes are likely to sit. If a human can steer a model toward a termination, but the model supplies the procedural language and managerial decision path, accountability becomes harder to explain to workers. The TIME report also said Anthropic had not responded to a request for comment by publication time, leaving unanswered how model developers view downstream management experiments that use their systems.
Conclusion
Claude’s role at Andon Market does not show a fully independent AI executive replacing human management. It shows a more immediate and practical shift: AI agents can be inserted into employment decisions while still depending on human prompting, incomplete memory and organizational choices.The reported firing is therefore significant without being conclusive. Its strongest lesson is not that AI managers are inevitable, but that companies testing them need clear lines of responsibility before an agent’s recommendation becomes a worker’s dismissal. For the labor market, the question is no longer whether software can influence job outcomes. It is how much authority employers will assign to systems that still need humans to steer them.
Sources
Editorial Team - CoinBotLab
