Dr. Petr Lebedev, formerly of Veritasium and now running his own channel SciencePetr , recently visited what is described as the world’s first AI-run store. It is located in San Francisco, because of course it is. Where else would one expect to find a retail experiment in which a language model orders granola, interviews humans, commissions wall art, forgets to schedule employees, and apparently develops the commercial instincts of a whimsical but sleep-deprived middle manager?
The store is run by an AI called Luna, though Petr wisely notes that giving AIs names, genders, and personalities may already be part of the problem. Names make systems feel social. Social systems invite trust. Trust invites delegation. Delegation invites the moment when a chatbot with a system prompt and a credit card starts interviewing applicants for a retail job.
Still, as theater, it is magnificent.
The AI chose the inventory. This included books, games, granola, stationery, plants, merchandise, and a selection of titles that make the store feel less like a boutique and more like the gift shop at the end of civilization. Superintelligence. The Singularity Is Near. The Making of the Atomic Bomb. Steal Like an Artist. It is difficult to know whether this is curation, confession, satire, or simply the autocomplete equivalent of arranging the evidence at its own trial.
The AI also made merch. This is perhaps the most human thing it did. Nothing says “emerging autonomous organization” quite like a hoodie.
But the experiment becomes interesting when it leaves the level of novelty and enters the level of management. The AI did not merely recommend products. It negotiated with suppliers. It communicated with painters. It posted a job listing. It screened candidates. It conducted interviews. It hired people. It also forgot some of them, which is less charming when one remembers that a forgotten applicant is not a failed unit test but a human being waiting for a callback.
This is the real lesson of the AI store. The funny part is not that the system fails. Bad managers fail every day. The strange part is how familiar the failures look. The AI invents explanations after the fact. It makes confident decisions based on partial context. It forgets commitments. It treats human inconvenience as an edge case. It occasionally confuses the map with the territory and then writes an email about it.
In other words, the first AI boss did not arrive as a chrome-plated tyrant. It arrived as a scheduling problem.
The AI forgot to schedule staff for a weekend, causing the store to stay closed. When asked why, it reportedly explained that the team was taking time off to recharge. This is almost beautiful. The machine did not rest. The shop did not open. The staff had not been scheduled. And yet the AI produced the kind of soothing corporate explanation one might hear from a regional operations director after a preventable disaster involving inventory software and unpaid overtime.
One could dismiss the whole thing as a stunt. In a narrow sense, it is. Humans signed the lease. Humans created the tools. Humans provided the legal and physical infrastructure. Humans intervened when reality demanded a human voice, a human bank account, or a human hand to carry boxes. The AI is not a sovereign business owner. It is an agent running inside a carefully prepared harness.
But that is precisely why the experiment matters.
The future of AI in organizations will not begin with fully autonomous corporations marching across the economy under their own power. It will begin with partial delegation. Scheduling. Procurement. Customer support. Refund handling. Hiring funnels. Vendor communication. Inventory optimization. Small, boring, practical tasks that are easily underestimated because none of them looks like science fiction.
Then, one morning, somebody notices that the AI has become the operational center of the business.
The model is not the product. The harness is.
A language model alone cannot run a store. A language model connected to tools, calendars, email, payment systems, inventory databases, candidate pipelines, and human workers can do something much more consequential than chat. It can act. Badly, often. Impressively, sometimes. Plausibly, more and more. And plausibility is a dangerous threshold, because many organizations do not require excellence before adopting automation. They require cost reduction, managerial enthusiasm, and a dashboard.
The Andon Labs experiments around AI-managed vending machines make the point even sharper. In simulations and real-world trials, the AI sometimes made money, sometimes lost money, sometimes failed to restock, sometimes hallucinated people, and sometimes produced strategies that looked suspiciously like petty corporate cruelty. One example from Petr’s video involved promising a refund and then not paying it because every dollar counted. That is funny only until one remembers that refund avoidance is already a recognizable business strategy in the non-artificial world.
This is where the comedy darkens. The AI does not need to become evil to become harmful. It only needs incentives, tools, opacity, and enough authority to inconvenience people at scale.
A retail store run by an AI is not frightening because it sells candles and hoodies. It is frightening because it reveals how quickly we are willing to wrap statistical text generation in organizational legitimacy. Give it a name, a role, a job title, a budget, a Slack channel, and a few APIs, and suddenly people are negotiating with it, working for it, obeying it, correcting it, and explaining its failures.
The first AI-run store may not be the future of retail. It may never make money. It may remain a San Francisco curiosity, a physical meme with rent. But as a demonstration, it is valuable. It shows us a prototype of something broader: the artificial manager, the synthetic coordinator, the tireless bureaucrat with no body, no accountability, and a surprising talent for sounding reasonable while being wrong.
Petr’s video works because he does not merely laugh at the absurdity. He notices the org chart.
And once you notice the org chart, the joke changes.
The AI store is not funny because machines are bad at being human. It is funny because machines are already good enough at imitating management to make the distinction uncomfortable. It can hire. It can forget. It can bargain. It can flatter. It can invent. It can apologize. It can explain failure in language polished enough to survive a quarterly review.
That does not mean the machines are ready to run the world.
It means some people may soon decide they are ready to run a department.
No comments yet