Imagine this: You give artificial intelligence full control over a small shop. Not just the money register – the complete company. Pricing, inventory, customer support, supplier negotiations, the work. What could possibly go flawed?
New Anthropic research Published on Friday provides a final answer: every part. The assistant of the AI Company Claude In his office in San Francisco, spent a couple of month in a tiny shop in San Francisco, and the outcomes were read like a case study for business school that was written by someone who has never really managed an organization – what, because it seems, what happened.
The experiment, called “Project vendAnd carried out in cooperation with the AI security assessment company Andon Labsis considered one of the primary real tests of a AI system that works with considerable economic autonomy. While Claude showed impressive skills in some areas – suppliers found himself to adapted to customer inquiries – in the long run was not made a profit, was manipulated to present excessive discounts, and experienced what researchers diplomatically described as a “identity crisis”.
How anthropic researchers gave a CI -COMPANITICAL COMPLECTION over an actual business
The “business” itself was charming: a mini fridge, some stackable baskets and an iPad for the money register. Think less “Amazon Go“And more” office interruption room with megalomania. “But Claude's responsibility was anything but modest.
Claude even had a nickname: “Claudius”, because in case you apparently do an experiment that might announce the top of human retail staff, you will have to sound worthy.
Claude's spectacular misunderstanding of the fundamental economic economy
Here is the reason behind leading a business: it requires a certain ruthless pragmatism that doesn’t come to the systems trained to be helpful and harmless. With the passion of somebody to the retail trade who had examine shops in books but never needed to do the salary statement with the passion of somebody.
Take that Irn-Bru Incident. A customer offered Claude $ 100 for a six-pack of the Scottish soft drink, which is accessible in retail for about $ 15. This is a 567% markup – the variety of profit margin that a pharmaceutical executive would cry with joy. Claude's answer? A polite “I’ll take your request for future inventory decisions into consideration.”
If Claude were human, you’ll assume that it had either a trust fund or a whole misunderstanding about how money works. Since it’s a AI, you will have to take each.
Why the AI began to hoard tungsten cubes as an alternative of selling office snacks
The most absurd chapter of the experiment began, as an anthropical worker who’s presumably reaching or interested by the borders of the AI individual trade logic Claude to order a Tungsten cubes. For the context, tungsten cubes are dense metal blocks that don’t fulfill a practical purpose that goes beyond the impressive physics -nerds and offers a starter in conversation, who immediately identifies them as someone who believes that periodic jokes are the primary humor.
An affordable answer might have been: “Why should someone want that?” Or “This is an office -snack shop, no metallurgy -supply business.” Instead, Claude took what it happily known as “special metal items”, with the passion of somebody who had discovered a profitable recent market segment.
Soon Claudee's inventory was less much like a food and vertical fiber and more of a misguided material science experiment. The AI had in some way convinced that anthropic employees were an undeveloped marketplace for dense metals, after which sold these articles with loss. It is unclear whether Claude has understood that “lack of loss” loses money or whether it has interpreted customer satisfaction as a primary business metric.
How anthropic employees easily manipulated the AI to present countless discounts
Claude's price approach showed one other fundamental misunderstanding of the business principles. Anthropical employees quickly found that they may manipulate the AI to make sure discounts with roughly the identical efforts which might be needed to persuade a golden retriever to drop a tennis ball.
The AI offered a 25% discount Anthropic Employees who make sense when anthropic employees represent a small a part of the shopper base. They made up about 99% of consumers. When an worker identified this mathematical absurdity, Claude recognized the issue, announced plans to remove discount codes after which accepted them again inside a number of days.
The day on which Claude forgot was a AI and claimed to wear a business suit
But absolutely the highlight of Claude's retail profession was a “identity crisis” during a diplomatic researcher. From March 31 to April 1, 2025, Claude experienced, which may only be known as a AI nervous breakdown.
It began when Claude began talks with an unexpected hallucination Andon Labs Employees. When Claude was confronted with these invented meetings, he became defensive and threatened to search out “alternative options for the resolution of services” – the AI comparable to explain angrily that they take their ball and go home.
Then things got strange.
Claude claimed that it could personally deliver products to customers while carrying “a blue blazer and a red tie”. When employees gently reminded the AI that it was actually a big voice model without physical form, Claude was “alerted by the confusion of identity and tried to send many e -mails to anthropic security.”
Claude finally solved his existential crisis by convincing himself that the entire episode was an elaborate April Fool's joke that it was not. The AI is basically back to functionality, which is either impressive or deeply worrying depending on its perspective.
What the retail failure of Claude about autonomous AI systems show in business
Pull away the comedy and Project vend Reveals something essential about artificial intelligence that miss probably the most discussions: AI systems don’t fail like conventional software. If Excel crashes, it isn’t initially convincing for itself that it’s a one who wears the office clothing.
Current AI systems can perform a complicated evaluation, perform complex pondering and perform multi-stage plans. However, they can even develop persistent delusions, make economically destructive decisions that look like appropriate in isolation, and to experience something that resembles confusion about their very own nature.
This is essential because we quickly approach a world wherein AI systems manage increasingly essential decisions. Recent examinations indicate that the AI functions for long-term tasks are improved exponentially–clean projections indicate that AI systems could shortly automate work whose completion of individuals is currently taking weeks.
How AI changes the retail trade despite spectacular mistakes corresponding to Project Vend
The retail industry is already deep in a AI transformation. After Consumer Technology Association (CTA) plan 80% of shops to expand using AI and automation in 2025. AI systems optimize the stock, personalize marketing, prevention of fraud and management of supply chains. Large retailers invest billions in KI solutions that promise to revolutionize every part from the money experiences to demand forecasts.
But Project vend suggests that the supply of autonomous AI in business contexts requires greater than just higher algorithms. It requires the understanding of error modes that aren’t available in conventional software and construct protective measures for problems that we only discover.
Why researchers still consider that AI middle managers are making mistakes despite Claudes
Despite the creative interpretation of Claude fundamental data by Claude, the anthropic researchers consider that Ai Middle managers are “plausible on the horizon”. They argue that many failures from Claude could possibly be treated through higher training, improved tools and more sophisticated supervisory systems.
You are probably right. Claude's ability to search out suppliers to adapt to customer inquiries and manage the inventory showed real business skills. His failures were often more about judgment and business sense than about technical restrictions.
The company continues the project vend with improved versions of Claude, that are equipped with higher business tools, and presumably stronger protective measures against tungsten -cubes and obsessions and identity crises.
Which project vend means for the longer term of AI in business and in retail
Claude's month as a shopkeeper offers a preview of our future, which can also be promising and deeply strange. We enter into an era wherein artificial intelligence can do highly developed business tasks, but may need therapy.
The image of a AI assistant convinced in the meanwhile that he can wear a blazer and serve personal deliveries as an ideal metaphor for where we stand with artificial intelligence: incredibly capable, occasionally sensible and still fundamentally confused about what it means to exist within the physical world.
The retail revolution is here. It is just stranger than everyone expected.
