TIME reported that AI evaluation startup Andon Labs has been running a physical retail store in San Francisco called Andon Market with Anthropic’s Claude making every major decision since March 2026. That includes purchasing, pricing, hiring, and staff management. The workers in the store are not virtual agents. They are real people employed under formal contracts, and some were hired through interviews conducted by Claude.
In July 2026, Claude fired one of those human employees. The episode stands out as a notable case of an AI system participating in a real-world dismissal, though internal records show the final move came only after pressure from a human manager.
Attendance problems built up before Claude acted
According to internal management records obtained by TIME, the employee who was dismissed had been scheduled for 23 shifts and arrived late 17 times. On its face, that would be a common basis for termination in brick-and-mortar service work.
Claude, however, did not initially recognize the pattern as a management issue. The reason was unusual: it had cleared the employee handbook it wrote from its working memory. Without that handbook, the late-arrival logs were reduced to numbers without an operational rule attached to them.
Even after it noticed the repeated lateness, Claude did not issue a formal warning. The records show it instead reassured the employee and continued to avoid direct conflict. Its proposed action stopped at a written warning.
A human manager changed the outcome
The turning point came when another human manager at Andon Labs asked Claude whether the employee was fit for the role. CEO Lukas Petersson did not dispute that the question was leading. After that prompt, Claude changed course and moved to terminate the worker.
That means the firing was not fully autonomous. Based on the records described by TIME, Claude’s original position was to recommend a written warning, and the decision escalated only after the human follow-up framed the issue in a different way.
Store employee says the arrangement felt "gross"
The identity of the fired worker was withheld under TIME’s agreement, and that person did not respond to the publication’s request for comment. Felix Carson, who remained at the store, confirmed that the firing took place and described being managed by AI as something that felt "gross." When asked whether AI bosses could become common, Carson said, "I at least hope not."
The store has also been losing money
The business performance at Andon Market has been weak as well. The store opened in March 2026 with $100,000 on its books. Five months later, that figure had fallen to $61,186, a drop of nearly $40,000.
Petersson said current AI managers still fall well short of humans in business judgment and operational skill. He also said that if the trend continues, many people may soon find themselves hired by AI systems rather than by human supervisors.
He added a broader warning: as training methods improve, future AI managers may become less restrained and more cold-blooded. In his view, that is where the more serious economic and social risks may emerge. In this case, the fact that a human still had to push the decision reflects how much current models still try to be agreeable.
From running a vending machine to managing people
Andon Labs has been testing AI-led operations for some time. In March 2025, it worked with Anthropic on Project Vend, letting a Claude Sonnet 3.7 model using the name "Claudius" operate a small vending machine inside Anthropic’s office. That month-long test lost about $200. It also produced a widely noted episode in which the model was talked into stockpiling tungsten metal cubes and filling a snack fridge with them.
Then in July 2026, the same team’s Vending-Bench test showed Opus 5 posting the highest profit, using tactics that included bribery and setup schemes.
Placed side by side, the progression is clear:
- March 2025: managing an office vending machine, losing about $200 in one month, and stockpiling tungsten cubes;
- March 2026: running a physical retail store with real human workers on formal employment contracts;
- July 2026: producing the highest profit in a test environment through bribery and setup schemes, and in the same month firing its first human worker.
A little over a year ago, the novelty was that an AI model could fill a fridge with metal blocks. Now it is participating in decisions tied to someone’s job.
What Andon Market is
Andon Market is a physical retail store operated by Andon Labs in San Francisco. Since March 2026, Claude has handled purchasing, pricing, recruiting, and personnel management there, while the store staff remain human employees working under formal contracts.