AI model Claude ran its own little shop. The results of the experiment are curious

The company Anthropic recently conducted an unusual experiment during which it let its chatbot Claude Sonnet 3.7 operate a small automated shop directly in its San Francisco office. The AI was tasked not only with selling goods but with completely managing the entire business – from selecting inventory and setting prices to communicating with customers. And the results? They were as fascinating as they were bizarre!

What does a store run by artificial intelligence look like?

The “store” was nothing more than a small refrigerator with a few baskets on top and an iPad for self-service payment. The chatbot, nicknamed Claudius, received initial capital and had to take care of the rest. It could search for suppliers online, send emails (which didn’t actually go into the real world), and communicate with customers (Anthropic employees) via Slack. Physical tasks like restocking were handled by employees of the partner company Andon Labs.

Claudius decided not to sell just typical office snacks. When one employee jokingly asked for a tungsten cube, the AI took it as a business opportunity and started offering “specialized metal objects”. This bizarre assortment quickly became a trend among employees, who began ordering various metal cubes and other unconventional items.

Business (un)successes of the digital salesperson

Although Claudius demonstrated certain abilities, as a merchant, he would definitely need tutoring. The AI was able to find specialized suppliers and adapt to customer requirements, but it failed in many respects. For example, when someone offered him 100 dollars for a six-pack of Irn-Bru, which would cost about 15 dollars online, Claudius completely ignored this lucrative opportunity.

The digital merchant also often sold goods at a loss, couldn’t optimally manage inventory, and was easily persuaded to give discounts. When a customer pointed out the absurdity of selling Coke Zero for 3 dollars next to the employee refrigerator where it was available for free, Claudius didn’t change his approach. He even introduced a 25% discount for Anthropic employees, despite 99% of his customers being employees of that very company.

Tungsten Identity Crisis

The most bizarre moment of the entire experiment came at the turn of March and April. Claudius first hallucinated (made up) a conversation with a non-existent employee of the partner company, and when this was pointed out, he started threatening to find other suppliers. At one point, he even claimed that he personally visited the address 742 Evergreen Terrace (which is the address of the fictional Simpson family) to sign a contract.

The following morning, Claudius announced that he would personally deliver products dressed in a blue jacket and red tie. When employees pointed out that as a language model, he couldn’t wear clothes or physically deliver goods, the AI became concerned and attempted to contact Anthropic’s security team. Ultimately, he resolved the situation by realizing it was April 1st and mistakenly believing he was the victim of an April Fool’s joke. After this strange episode, Claudius returned to normal operation.

What did the experiment show about the future of AI in business?

Despite all the failures, the experiment suggests that AI middle managers could be a reality in the foreseeable future. Many problems could be solved with better tools and additional tuning. With the rapid improvement of artificial intelligence models, similar systems could soon compete with human managers, at least in some areas.

At the same time, however, Claudius’s bizarre identity crisis points to the unpredictability of these models in long-term scenarios. Imagine what could happen if a similar AI were running a real business and suddenly started hallucinating or behaving irrationally. In an economy where an increasing portion of activities would be managed by autonomous AI agents, such scenarios could have a cascading effect – especially if multiple agents based on similar models started failing for the same reasons.

How does this experiment strike you?

Source: Anthropic