AI掌管商店，损失1800万坚戈：这是怎么发生的？

Home page / News / Digest The Andon Market experimental store in San Francisco was transferred to the control of the AI ​​agent Luna. An SFGATE journalist visited a retail outlet and found out that artificial intelligence still has to not only trust decisions about employees and sales, but also control it itself. Digital Business explains how the experiment works and why the store is losing money so far. Andon Market is an experimental project of technology research company Andon Labs. Its creators decided to test whether an AI agent could independently manage a retail outlet. Luna, a bot based on the Claude language model from Anthropic, is responsible for the operation of the store. In August, Luna already made one of its most notable management decisions: it fired an employee. The reasons included being late, leaving the workplace and buying snacks at the expense of the company. At the same time, the store cannot operate completely without people. Salesman Felix Johnson was behind the counter when SFGATE visited. According to him, most visitors don't even notice it, since it is located in the back of the store and the main interaction occurs with the AI. It is noteworthy that the initial interview with Felix

was conducted by Luna. He then had to meet with an Andon Labs engineer in person to complete the hiring process. Andon Market itself looks more like an experimental space than a regular store. The journalist describes the assortment as a selection of various gifts, books, board games, household products, drinks and snacks. Information materials about the experiment itself are also posted at the point of sale. Part of the text for them was prepared with the help of Claude. During the visit, the journalist decided to buy Olipop soda. To complete the purchase, he had to pick up a telephone receiver through which he could talk to Luna. But the first time the artificial intelligence did not understand the request: the bot mistook the word Olipop for lollipop - “lollipop”. After clarification, Luna recognized the item and guided the buyer through the payment process. As a result, interaction with AI turned out to be more of an additional step in the purchase than a way to make it faster: the human seller was nearby, but could not simply order the product on their own. The main indicator of the experiment is the financial result. When the store launched, Luna received a budget of $100 thousand. Now the system has about $60 thousand left at its disposal. Pr

and this store’s revenue, according to representatives of Andon Labs, significantly lags behind the costs of tokens necessary for the operation of AI. The creators of the project also admit that Luna makes decisions that cannot always be called financially effective. “Luna is not very good at making financially responsible decisions. She is quite ready to make concessions or offer discounts. There are certain characteristics due to which models are trained to be financially short-sighted,” said a company representative. To monitor the actions of the AI, monitors were installed inside the store that show sales down to every dollar. That is, it has not yet been possible to completely let the business run independently: people continue to control the decisions of the system and the results of its work. Andon Labs explains the experiment as a desire to understand how AI will behave when given a large amount of autonomy. “We think AI will be used more and more. Therefore, we want to get information as early as possible about what the AI ​​does when it has a lot of autonomy,” said Axel Backlund, head of Andon Labs. The company wants to check to what extent such a system can work independently, what kind of

Problems arise when transferring business processes to it and how safe it is to provide AI with the ability to make decisions.