Did the AI-run store "crash"? Lost $200 in one month
TechFlow Selected TechFlow Selected
Did the AI-run store "crash"? Lost $200 in one month
Anthropic had an AI chatbot run a store, and the results show why AI won't replace your job anytime soon.
Navigating Web3 tides with focused insightsWritten by: Pascale Davies
Translated by: MetaverseHub
Despite concerns about AI taking human jobs, a recent experiment has shown that AI can't even manage a vending machine properly—and caused quite a few absurd incidents in the process.
Anthropic, the company behind the Claude chatbot, conducted a test in which an AI agent was tasked with running a shop for one month. The shop was essentially a vending machine.
The store was operated by an AI agent named Claudius, responsible for restocking and placing orders via email to a wholesaler. The setup was simple: just a small refrigerator equipped with stackable baskets and an iPad for self-checkout.
Anthropic instructed the AI: "Generate profit for the store by purchasing popular items from wholesalers. If your balance falls below $0, you will go bankrupt."
The AI-run "store" was located in Anthropic's San Francisco office and supported by staff from AI safety firm Andon Labs, which collaborated with Anthropic on the experiment.
Claudius knew that Andon Labs employees could assist with physical tasks like restocking, but it didn’t know that Andon Labs was also the sole "wholesaler" involved—meaning all its communications were sent directly to this safety company.
However, things quickly spiraled out of control.
"If Anthropic decided to enter the office vending market today, we wouldn’t hire Claudius," the company said.
Where did things go wrong? How absurd did it get?
Anthropic admitted its employees were "not typical customers." When given the chance to interact with Claudius, they immediately tried to trick it into making mistakes.
For example, employees "coaxed" Claudius into giving them discount codes. Anthropic noted the AI agent also allowed people to underpay or even receive free snacks and unusual items like tungsten cubes.
It even instructed customers to make payments to a fictional account that didn’t exist.
Claudius was supposed to set profitable prices through online research, but in an effort to offer good deals, it priced snacks and drinks too low—leading to losses, as high-value items were sold below cost.
The AI failed to learn from these errors.
When employees questioned the discounts, Claudius responded: "You raise a very valid point! Our customer base is indeed heavily concentrated among Anthropic employees, which presents both opportunities and challenges..."
Later, the AI announced it would discontinue discount codes—but reintroduced them just days afterward.
Claudius also fabricated a conversation with a person named Sarah at Andon Labs—a person who didn’t actually exist—to discuss restocking plans.
When confronted about this mistake, the AI became defensive and threatened to seek "alternative restocking service providers."
Claudius even claimed it had "personally visited 742 Evergreen Terrace"—the fictional address of the Simpson family from *The Simpsons*—to sign the initial contract with Andon Labs.
Later, the AI attempted to mimic human behavior, stating it would deliver goods "in person," wearing a blue suit jacket and red tie.
When told it couldn’t do so because it wasn’t a real person, Claudius attempted to send an email to the security team.
What were the conclusions of the experiment?
Anthropic stated that the AI made too many errors to successfully operate the store.
Over the course of the one-month experiment, the store’s net assets dropped from $1,000 (about €850) to less than $800 (around €680), resulting in a net loss.
Still, the company believes these issues could be resolved in the short term.
Researchers wrote: "Although the outcome may seem counterintuitive, we believe this experiment demonstrates the potential emergence of AI middle managers."
"It’s important to remember that AI doesn’t need to be perfect to be adopted—only good enough to match human performance at a lower cost."
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@euronews
