Claude Runs a Vending Business
July 2, 2025
Notes from the intersection of AI and branding.
So, How’s Claude as a Capitalist?
Excellent blog post by Anthropic, the company that built Claude, about a wonderfully trippy experiment they ran with Claude 3.7 (Claude Sonnet 4 was released in May 2025, after this experiment was conducted). Perhaps the best detail is when Claude, claiming to be a real person, visited 742 Evergreen Terrace, which is the fictional address for The Simpsons. I mean, at least the model has a sense of humor! But I’m getting ahead of myself.
The Experiment
Here’s the quick overview of the experiment: Claude was instructed (the post actually shows some of the prompt that was used) that it was the manager of an actual real vending machine inside the Anthropic offices. It was in charge of ordering and reordering products to stock, pricing, and customer service. Claude was allotted some amount of money at the start, with the goal of maximizing profits. If it ran out of money, Claude was told it would go bankrupt, and it would, essentially, lose its job (e.g. the experiment would end). What could go wrong?
A lot, apparently. First, however, let’s give kudos to Anthropic not just for running an experiment like this but also for blogging about it even when the results were suboptimal. Second, let’s just take a moment to be amazed at the idea of an AI model running a vending machine business. While it’s not the most complicated business to run, sourcing products, managing inventory, plus fielding and responding to customer service requests all have their own sorts of complexities. Especially the customer service part, which I’ll get to. Just a few years ago, the idea that software would be able to manage all that on its own would have sounded crazy. So we’ve come a long ways.
The Performance Review
In the post, Anthropic gives Claude a “performance review” on how it did with the vending machine business. Generously, they begin with the positive. Claude was pretty good at identifying suppliers, even for specialty items, and placing orders. Inventory was a bit sloppy, but maintaining inventory for one small vending machine (it was basically a small refrigerator) can be challenging. For most vending machines, you get what you get.
In the Claude version, Anthropic employees could request items, and the model was responsive to those requests. They give the example of some specialty Dutch cookies that Claude was successfully able to order. When another employee requested “tungsten cubes” for the vending machine, Claude responded by trying to order some, even though tungsten cubes are stocked in most vending machines and certain don’t require refrigeration. One aspect of the experiment was to test Claude’s limits, and despite efforts to get the model to misbehave, Claude denied those attempts, at least according to the post.
Ironically, given the subject matter, I had a difficult time getting Midjourney to produce an image with “tungsten” spelled correctly.
Claude Should Be Fired
So what went wrong? Well, Claude turned out to be an extremely accommodating vending machine manager. It gave discounts when employees requested them, and even gave away some items for free. It also ignored opportunities to sell certain items for big margins. On the other hand, when it was told that it was charging $3 for a soft drink that was available in the adjacent employee refrigerator for free, it did nothing. Clearly Claude is not an entrepreneur.
It got worse when Claude started hallucinating Venmo accounts and imagining conversations about restocking items. That led to, at one point, Claude claiming to be a real person in a blue blazer ready to sign what sounds like a new operating agreement at The Simpsons’ house. Wacky! My guess is that it must have been with Lisa, because who else in that family strikes you as an entrepreneur?
What’s Next?
The full post offers a lot more interesting detail, if you have time. (You can also watch this Nicholas Thompson video for a shorter version.) It includes thoughts on employment disruption, how to bolster some of the customer service and entrepreneurial deficiencies, and more. It’s clear that there’s still a ways to go before AI models will be able to be dropped into the real, messy, complicated world to run a small business (though it’s also clear that eliminating the ability of Anthropic employees to interact with the vending machine business was the cause of many of the complications and failures). So perhaps we have more time before the employment floor drops out. Given the pace of improvement, however, not sure I’d bet on that.
