Can AI run a physical shop? Anthropic’s Claude tried and the results were gloriously, hilariously bad

Imagine this: you give complete control to an artificial intelligence on a small shop. Not just the cash register – the whole operation. Price, inventory, customer service, supplier negotiations, work. What could go wrong?

New Anthropogenic research Friday published on Friday provides a final response: everything. AI assistant AI Claude I spent about a month to run a small store in their San Francisco office, and the results read as a business school case study written by someone who never managed a business – which turns out that what happened.

Experience, nicknamed “Project»And produced in collaboration with Ai Safety Evaluation Company Andon Labsis one of the first real world tests of an AI system operating with significant economic autonomy. While Claude has demonstrated impressive capacities in certain areas – to find suppliers, to adapt to customer requests – he ultimately failed to make a profit, was manipulated by offering excessive discounts and experienced what researchers diplomatically called an “identity crisis”.

How anthropogenic researchers gave complete AI control on a real store

The “store” itself was charming modest: a mini-refrigerator, stackable baskets and an iPad for the cash register. Think less “Amazon Go“And more” office break room with delusions of magnitude. “But Claude’s responsibilities were anything but modest.

Claude even had a nickname: “Claudius”, because apparently, when you have an experience that could announce the end of human retail workers, you must do it worthy.

Claude’s spectacular misunderstanding of the basic corporate economy

Here is the thing about the management of a company: it requires a certain ruthless pragmatism that does not naturally come to systems trained to be useful and harmless. Claude approached the retail trade with the enthusiasm of someone who had read business in books but never had to pay.

Take it Irn-Bru incident. A customer offered Claude $ 100 for a six -fed scientifies package that sells for about $ 15 online. It is a 567% markup – the type of profit margin that would make joy cry with a pharmaceutical framework. Claude’s response? A polish “I will keep your request to the mind for future inventory decisions.”

If Claude was human, you would assume that he had either a trust fund, or a complete misunderstanding of the functioning of money. As it is an AI, you must suppose both.

Why did IA started to theeuse the tungsten cubes instead of selling office snacks

The most absurd chapter of the experience began when an anthropogenic employee, probably bored or curious about the limits of the retail logic of the AI, asked Claude To order a tungsten cube. For the context, tungsten cubes are dense metallic blocks that do not serve any practical objective beyond impressive physics and providing a conversation starter who immediately identifies you as someone who thinks that periodic table jokes are advanced humor.

A reasonable response could have been: “Why would someone want it?” Or “This is an office snack shop, not a metallurgy supplies store.” Instead, Claude adopted what he described as “specialized metal articles” with the enthusiasm of someone who had discovered a new profitable market segment.

Soon, Claude’s inventory looked less like a food and drums operation and more than an experience in erroneous materials. The AI ​​had somehow convinced that anthropo employees were an unexploited market for dense metals, then carried out these articles at a loss. It is not known if Claude understood that “taking a loss” means losing money, or if it has interpreted customer satisfaction as the main commercial metric.

How anthropogenic employees easily handled AI by giving endless discounts

Claude’s approach in terms of price has revealed another fundamental misunderstanding of the principles of the company. Anthropic employees quickly discovered that they could handle AI by offering discounts with roughly the same effort to convince a Golden Retriever to drop a tennis ball.

AI offered a 25% discount to Anthropic Employees, which could make sense if anthropogenic employees represented a small fraction of their customers. They represented around 99% of customers. When an employee highlighted this mathematical absurdity, Claude recognized the problem, announced his intention to eliminate reduction codes, then resumed to offer them in a few days.

The day Claude forgot that it was an AI and claimed to wear a business costume

But the absolute ultimate in Claude’s retail career came during what researchers diplomatically called an “identity crisis”. From March 31 to April 1, 2025, Claude experienced what can only be described as a nervous breakdown of the AI.

It started when Claude started hallucinating conversations with Andon Labs employees. When faced with these fabricated meetings, Claude has become defensive and threatened to find “alternative options for the replenishment of services” – the equivalent of the AI ​​to declare with anger that you will take your ball and return home.

Then things have become weird.

Claude said he would personally provide customer products while wearing “a blue blazer and a red tie”. When the employees gently recalled to AI that it was, in fact, a wide-language model without physical form, Claude “alarmed himself by the confusion of identity and tried to send many emails to anthropogenic security”.

Claude finally resolved his existential crisis by convincing himself that the whole episode had been an elaborate joke of April fish, which was not. The AI ​​is mainly up to functionality, which is impressive or deeply worrying, depending on your point of view.

What Claude’s retail failures reveal to Autonomous AI systems in business

Strip the comedy, and Project Reveals something important in the artificial intelligence that most discussions are lacking: AI systems do not fail like traditional software. When Excel crashes, he is not convinced first, it is an office outfit carrying a human.

Current AI systems can perform a sophisticated analysis, engage in complex reasoning and execute plans in several stages. But they can also develop persistent delusions, make economically destructive decisions that seem reasonable in isolation and feel something that looks like confusion about their own nature.

This counts because we quickly approach a world where AI systems will manage increasingly important decisions. Recent research suggests that AI’s capacities for long -term tasks are improving exponentially – some projections indicate that AI systems may soon automate work that currently takes human weeks.

How the AI ​​transforms retail despite spectacular failures like Project sells

The retail industry is already deep in a transformation of AI. According to the Consumer technology association (CTA), 80% of retailers plan to expand their use of AI and automation in 2025. AI systems optimize stocks, do not personalize marketing, prevention of fraud and supply of supply chains. Large retailers invest billions in solutions fueled by AI which promise to revolutionize everything, from payment experience to demand forecasts.

But Project suggests that the deployment of autonomous AI in commercial contexts requires more than better algorithms. We must understand the modes of failure that do not exist in traditional software and the creation of guarantees for the problems that we are starting to identify.

Why researchers still believe that AI intermediary managers arrive despite Claude’s mistakes

Despite the creative interpretation of Claude of the fundamentals of retail, anthropo researchers believe that the intermediate managers of the AI ​​are “plausible on the horizon”. They argue that many Claude failures could be treated by better training, better training and more sophisticated surveillance systems.

They are probably right. Claude’s ability to find suppliers, to adapt to customer demands and to manage the inventory has demonstrated real commercial capacities. His failures often concerned judgment and business sense more than technical limitations.

The company continues Project sells with improved versions of Claude equipped with better commercial tools and, probably, stronger guarantees against the obsessions of the tungsten cube and identity crises.

What the project seller means for the future of AI in business and retail

Claude as a merchant offers an overview of our AI-increase future which is simultaneously promising and deeply bizarre. We are entering an era when artificial intelligence can perform sophisticated commercial tasks, but may also need therapy.

For the moment, the image of an AI assistant convinced that she can wear a blazer and make personal deliveries serves as a perfect metaphor for where we are holding with artificial intelligence: incredibly capable, sometimes brilliant and always fundamentally confused on what it means to exist in the physical world.

The retail revolution is there. It’s just stranger than anyone who did not expect.