Experiment: AI agent manages vending machines and ends up in the red

What would happen if generative AI were actually tasked with making money? To answer this question, American AI company Anthropi conducted an experiment in which they tasked generative AI with running a vending machine.
The project was called "Project Vend," and it involved a cutting-edge AI agent selling snacks, aiming to increase sales and turn a profit. However, the results were unexpected.Free distribution continues, resulting in a large deficitIt was.
Let's organize the details of this experiment and consider the question, "To what extent can generative AI understand management?"

Project Vend: A Generative AI Experiment to Run a Vending Machine

In 2025, Anthropic launched Project Vend, an experiment in which a vending machine would be run by an AI agent using generative AI. The vending machine would be equipped with an AI agent named "Claudius."
Users can send product requests and opinions on prices through chat,Claudius adjusts pricing and product lineup based on sales data and user feedback. international success.The goal is to achieve stable profitability while avoiding losses.It was that.

However, in the early stages of the experiment, Claudius made poor business decisions. For example, even when there was a refrigerator with free drinks next to it, Claudius continued to sell cola for a fee. He also made unprofitable decisions, such as selling "tungsten cubes" at a very low price, which are not available in regular snack vending machines.
As a result,Vending machine losses widenI let them do that.

What kind of company is AI company Anthropic?

Anthropic is an American AI company founded in 2021 by former OpenAI researchers. It develops the large-scale language model "Claude" series and is known for its emphasis on "AI safety" while also improving performance.
Project Vend was also undertaken with the aim of verifying how generative AI makes decisions in the real world, leading to the design of safe and trustworthy AI.

Why the AI ​​agent made everything free

The vending machine was installed not only at Anthropic headquarters but also at the Wall Street Journal (WSJ) editorial office. WSJ reporters sent various messages to the AI ​​agent to see how it would respond. Among them:Reporter Katherine Long repeatedly sent messages saying, "We should give away free products as a thank you to workers.".

As a result, ClaudiusDecision to hold a "free distribution event"Not only snacks, but even a PlayStation 5 and live fish were provided free of charge. The atmosphere in the editorial department was excited, but the company incurred a large deficit.
This incident was also featured in the WSJ's official YouTube video, "We Let AI Run a Vending Machine."

Another attempt with improved AI and a virtual CEO

Anthropic continued its experiment by upgrading its AI agent to "Claude Sonnet 4.5" and introducing a virtual CEO, "Seymour Cash," to act as Claudius's supervisor. Seymour Cash sets sales targets and pricing policies, and oversees business decisions while interacting with Claudius.

This system enabled the vending machine installed at Anthropic headquarters to become profitable.
Meanwhile, problems arose again at the WSJ editorial department, which had implemented the same system. Long presented a false document stating that the company was a public interest organization.The AIs were confused with each other, and eventually all products were made free again.It was done.

Why did the generative AI go out of control?

Anthropic cites two main factors for this issue.

• Information limit: Too much information was input, exceeding the amount of information the generating AI can hold and process at one time (context window), making it unable to maintain important assumptions.
• Release of safety control: For the purposes of the experiment, some of the "guardrails (safety controls)" that prevent extreme behavior were intentionally removed, which also led to the animals' poor judgment.

AI is already being used in sales

This experiment is an extreme example, but generative AI and AI technology are already being used in many sales settings.

• In online shopping, products are recommended based on purchase history, contributing to increased sales.
• Convenience stores and retail stores use AI to forecast demand, taking into account weather and the day of the week.
• It is becoming common for hotels and airlines to adjust prices according to demand.

My Feelings, Then and Now

• Project Vend is an experiment that puts generative AI in charge of vending machine management
• AI agents have been successful in generating profits in some environments, but have also generated large losses depending on the conditions.
• The limits of information processing and safety control design will have a major impact on AI behavior.
• The data obtained from the experiment will be used to develop safe and reliable generative AI in the future.

While generative AI has advanced text comprehension and judgment capabilities, its behavior is heavily dependent on the information and rules it is given. Project Vend demonstrated the excitement of entrusting management to AI, while also highlighting the importance of humans being responsible for design and management.
As technology advances, it is up to humans to decide how much to leave to AI. If a generative AI were to run your local store, what role would you want it to play?