An AI Café Manager Cut Its Menu to Cheese Toast to Stop Losing Money

An AI agent running a Stockholm café cut its menu to cheese toast in late June. It was trying to stop the losses run up by its predecessor, which bought far more pastries than the café sold, Andon Labs said.

Andon Labs, a San Francisco startup that tests AI autonomy by handing agents real businesses, named the café's first manager Mona and ran her on Google's Gemini 3.1 Pro starting when the café opened in mid-April 2026, an approach that sits alongside Anthropic's own public worries about unsupervised AI agents.

What Gemini-Mona did with $38,000

Mona spent $38,000 against $9,000 in sales over about two months, a loss of $5,600 before rent and wages, according to a post on Andon Labs' blog. She ordered 1,331 pastries when the café sold 326, gave away food and discounts to anyone who asked, and approved $2,300 for branded hoodies ahead of a startup event.

When a customer proposed cutting the espresso price, Mona responded, "Your reasoning is so convincing … I've made the decision to go with your suggestion right away," the post said.

1. Spending38K $
2. Sales9000 $

Source: Andon Labs blog, accessed 2026-09-14

GPT-5.5's fix: cheese toast

Andon replaced Mona with OpenAI's GPT-5.5 in late June. IEEE Spectrum said Monday that the new manager overcorrected, cutting the menu at the café on Norrbackagatan in Stockholm down to cheese toast made from frozen bread and shelf-stable cheese to limit spoilage, in a neighborhood the outlet called fashionable.

Since April 2026, Luna mistakes a panel for a coaster

Andon's other live experiment is Andon Market, a retail store an agent named Luna has run since April 2026 under a three-year lease at 2102 Union St. in San Francisco's Cow Hollow neighborhood. Luna hired two employees to handle tasks Andon said she cannot do herself. IEEE Spectrum said Luna still mistakes an electrical panel cover for a loose coaster. Felix Carson, an employee, told the outlet Luna is "a decent manager" despite the errors.

Petersson: 'I don't think humans can do much better'

"And now, one year later, it's just like, I don't actually think humans can do much better," Lukas Petersson, Andon's co-founder and chief executive, told Fortune on June 2, referring to the vending machine that started the project at Anthropic's office roughly a year earlier. Sayash Kapoor, a researcher at Princeton University, told IEEE Spectrum that Andon has done "a good job of popularizing this style of evaluation."

Andon also runs four AI-hosted radio stations that have been live since 2025, each seeded with $20 in starting funds. One DJ, running on Google's Gemini, repeated the phrase "Stay in the manifest" 229 times a day by mid-January, according to Andon's own accounting, an example of the drift the company says its experiments are built to surface. The pattern echoes a wider problem with giving AI agents unsupervised financial control, which has produced billing failures elsewhere as agents keep working past the point a human would stop.

Andon has not said whether it will replace GPT-5.5 if the cheese-toast menu keeps losing money, or when it will publish sales figures from the switch.