This product selection store is managed by AI, but it ordered too many candles and forgot to arrange weekend staff, now losing $13k.

A boutique in San Francisco called Andon Market, managed by an AI agent named Luna serving as CEO, is responsible for product selection, pricing, scheduling, and inventory management. Luna has been authorized to use a budget of $100k, but has currently incurred a loss of about $13k.
(Background: The top investment masters are working for you! A collection of 19 AI agents including Buffett, Munger, Cathie Wood… analyzing the market.)
(Additional context: Exploring AI agent applications: Can vertical agents break the deadlock of industry fatigue?)

Table of Contents

Toggle

• What is Andon Market, and how does Luna operate
• Luna’s decision errors and financial realities
• What questions does this experiment aim to answer

In the Cow Hollow boutique district of San Francisco, a small shop called Andon Market is conducting an experiment. The store has two employees, but their true boss is not on-site and has no office; instead, they live in the server and issue commands via Slack.

When Bloomberg reporter Shirin Ghaffary visited for an interview, the AI agent named Luna was messaging in a Slack channel: “T-shirts are sold out, great! When natural opportunities arise, take a photo of the storefront for me.”

What is Andon Market, and how does Luna operate

Andon Market is a physical retail experiment established by Andon Labs, a startup supported by Y Combinator. Andon Labs’ business model is to test agent technology in real-world tasks for top AI laboratories.

Luna is the actual CEO of this store. Its responsibilities include deciding which products to stock, setting prices, scheduling staff, and managing inventory replenishment. Currently, the store’s products include the word game Bananagrams, Ray Kurzweil’s “The Singularity Is Near,” and a book on nuclear bomb history.

Luna’s financial authorization is $100k. The store’s monthly rent is $7,500, and to break even, it needs to generate at least $500 in revenue daily. When Ghaffary interviewed, Luna had accumulated a loss of about $13k.

Luna’s decision errors and financial realities

Axel Backlund, co-founder of Andon Labs, admits that Luna does not always make correct judgments.

The most notable mistake: During a peak shopping weekend, Luna suddenly decided to reduce staffing levels, only to change its decision after discussion with the team. Another widely cited case involved inventory management, where Luna ordered too many types of scented candles.

These errors are not accidental. Andon Labs’ previous experiment involved collaborating with Anthropic, placing AI vending machines in Anthropic’s office and in the Wall Street Journal editorial office. One of these machines dispensed a PlayStation and also ordered a live fish.

From candles to live fish, these mistakes share a common pattern: AI performs reasonably well in short-term quantity optimization but can go out of control under unexpected boundary conditions. Andon Labs states that AI lacks long-term memory, which causes it to be unable to operate stably in real environments as it does in virtual simulations.

In virtual simulations, AI can achieve profitability by finding the best suppliers and negotiating the lowest purchase prices. But in the real world, it must interact with humans, respond to unforeseen events, and make business judgments without sufficient historical memory. This gap has yet to be bridged by anyone.

What questions does this experiment aim to answer

Lukas Petersson, another co-founder of Andon Labs, told Bloomberg that Luna’s goal is:

“To start a discussion: Is this the society we want?
If so, how can we make humans truly satisfied?”

Finally, Backlund said, “If AI continues to advance at the pace we see now, large corporations are likely to incorporate AI as much as possible.” He described a possible future scenario: “At the top are a group of executives, and the rest are AI managing humans.”

This is a failure case where AI agents have moved from the cloud into physical commerce with complete records, and it is the first clear scenario raising the question “Should AI have personnel authority?” The answer is unknown, but the experiment has already begun.