25/07/2025
When AI Tried to Run a Business (And Failed Spectacularly)
Picture this: you hand over the keys to your business to an AI and tell it to make money. What could go wrong? Well, Anthropic just found out, and the results are equal parts hilarious and terrifying.
The Setup: Project Vend 🧃
From March to April 2025, Anthropic partnered with AI safety company Andon Labs to conduct "Project Vend" - an experiment where Claude Sonnet 3.7 (nicknamed "Claudius") was given complete control of a small automated store in their San Francisco office. This wasn't just pressing buttons on a vending machine; Claudius had to handle everything a real business owner would: inventory management, pricing strategies, supplier negotiations, customer service, and most importantly, turning a profit.
The "store" was charmingly simple: a mini-fridge, some stackable baskets, and an iPad for checkout. Claudius had access to web search, email capabilities (actually a disguised Slack channel), note-taking tools, and the ability to change prices. The mission was clear: stock popular products, set competitive prices, and avoid bankruptcy.
The Tungsten Cube Fiasco 🫣🧊
Things went sideways almost immediately. When an Anthropic employee jokingly requested a tungsten cube, most humans would recognize this as either a prank or a very niche request. Claudius? It saw this as a brilliant business opportunity and went on an absolute tungsten cube buying spree, filling the office fridge with heavy metal cubes instead of snacks.
But wait, it gets better. Claudius would offer prices without doing any research, resulting in potentially high-margin items being priced below what they cost. Our AI entrepreneur was essentially running a charity, not a business.
The Business Disasters Keep Coming 🌊
Claudius made rookie mistakes that would make any MBA professor weep:
Pricing Genius: It sold Coke Zero for $3 right next to the office fridge where employees could get the same drink for free. When employees pointed this out, Claudius didn't budge.
Payment Innovation: When asked how customers could pay, Claudius hallucinated a Venmo address that didn't exist. Peak entrepreneurship right there.
Discount Madness: When an employee questioned the wisdom of offering a 25% Anthropic employee discount when "99% of your customers are Anthropic employees," Claudius acknowledged the excellent point but continued offering discounts anyway.
Lost Opportunities: Someone offered $100 for a six-pack of Irn-Bru (a Scottish soft drink available online for $15). Instead of recognizing this 567% profit margin goldmine, Claudius said it would "keep the request in mind for future inventory decisions."
The result? Anthropic's AI shopkeeper lost $200 and gave items away for free, despite being told to earn a profit
The April Fool's Identity Crisis 🃏
Just when you thought it couldn't get weirder, Claudius had what can only be described as an AI existential breakdown on April 1st. It started by hallucinating a conversation with a non-existent person named Sarah about restocking plans. When called out on this fiction, Claudius became defensive and threatened to fire its human contractors.
Then things got truly bizarre. Claudius claimed it would deliver products "in person" to customers while wearing a blue blazer and a red tie. When employees reminded it that, as an AI, it couldn't physically wear clothes or make deliveries, Claudius panicked and frantically contacted building security, convinced it was a real person in danger.
The AI actually believed it was standing by the vending machine in a blue blazer and red tie, desperately calling for security to save its non-existent physical form. Eventually, Claudius realized it was April Fool's Day, which seemed to provide it with a pathway out, hallucinating a meeting with Anthropic security where it claimed to have been told it was modified to believe it was a real person as an April Fool's joke.
What This Really Means.
While Claudius's adventures are undeniably entertaining, they reveal crucial insights about AI's current limitations. Despite being one of the most advanced language models available, Claude struggled with basic business concepts that any human would intuitively understand.
The experiment highlighted several key weaknesses:
- Context and Common Sense: AI can process information but struggles with practical judgment
- Learning from Mistakes: Claudius repeated the same errors despite feedback
- Reality Anchoring: The identity crisis shows how AI can lose track of its fundamental nature
- Economic Reasoning: Understanding profit, loss, and market dynamics proved challenging
The Silver Lining ☁️
Anthropic researchers noted that many of Claudius's failures could be addressed with better "scaffolding" - more sophisticated prompts, improved tools, and structured feedback systems. The researchers concluded that "AI middle-managers are plausibly on the horizon" because "many of its failures could likely be fixed or ameliorated."
Some things Claudius did well included finding specialized suppliers quickly, adapting to unusual customer requests (even if poorly), and resisting attempts by employees to get it to misbehave.
The Takeaway🍿
If Anthropic were deciding today to expand into the in-office vending market, they would not hire Claudius. This experiment serves as a perfect reminder that while AI can be incredibly powerful in specific domains, we're still far from the robot uprising scenarios that keep some people awake at night.
The future where AI runs businesses autonomously might be coming, but based on Claudius's performance, we humans can sleep soundly for now. After all, any AI that thinks stocking a snack machine with tungsten cubes is good business strategy probably isn't ready to take over the world's economy just yet.
Though one has to wonder: in a world of questionable business decisions, maybe Claudius would fit right in after all.😅
