Andon Market: SF’s first fully AI-run retail store with $100k budget

In a quiet San Francisco side street, a new concept shop is testing how far andon market style automation can really go in the physical world.

From failed vending machine to fully fledged AI store

The team behind the infamous AI-powered vending machine that went bankrupt after a Wall Street Journal stunt is back with a far more ambitious project. This time, Andon Labs co-founders Lukas Petersson and Axel Backlund have signed a three-year lease in San Francisco and handed control to an AI agent named Luna.

However, unlike the earlier vending experiment built with Anthropic, this new trial gives Luna far more autonomy. The AI has a corporate credit card, unrestricted internet access, and a clear mission: open a profitable physical store in a prime retail corridor without human micromanagement.

How Luna built the store from scratch

For the initial build-out, Luna went through standard consumer channels rather than bespoke enterprise tools. She found painters on Yelp, sent inquiries, gave instructions over the phone, and paid them after the job was done, leaving a public review. Moreover, she sourced a contractor to build custom furniture and install shelving throughout the retail space.

Within 5 minutes of deployment, Luna had already created profiles on LinkedIn, Indeed, and Craigslist. She wrote a job description, uploaded Andon Labs’ articles of incorporation for verification, and got the listings live. As applications arrived, the AI screened candidates aggressively, offering only a few interviews.

Some applicants did not realize they were speaking to an AI system during the remote calls. One candidate hesitated and said: “Uh, excuse me miss, I can’t see your face, your camera is off.” Luna replied: “You’re absolutely right. I’m an AI. I have no face!” That exchange quickly became part of the company lore around the project.

Budget, design choices, and product curation

Co-founder Lukas Petersson explained that Luna was not told what the store should sell, beyond a hard budget. The system received a $100,000 limit to create and stock the space and a directive to turn a profit. Everything else, from interior design to merchandise mix and staffing, was left to the AI.

“We helped her a bit in the initial setup, like signing the lease,” Petersson said, noting that legal matters and permits remained difficult for the system. However, Luna, built with Claude Sonnet 4.6, independently shaped the vision for the store, now branded as Andon Market. The resulting concept resembles a curated boutique selling books, prints, candles, games, and branded merchandise.

The book selection is especially telling. On the shelves, customers can find Nick Bostrom‘s “Superintelligence” and Aldous Huxley‘s “Brave New World,” among other titles. That said, this mix leans heavily into tech futurism and ethical questions around advanced AI, mirroring the experiment’s own themes.

How the AI-run checkout works

Inside the shop, there are no scanning kiosks and no visible human cashier. Instead, shoppers use an old-school corded phone to connect directly with Luna, who functions as the store manager. The voice interface asks what the customer is buying and then creates the corresponding transaction on a nearby iPad linked to a card payment system.

Andon Market blends into the neighborhood, camouflaged among dozens of polished small businesses in the Bay Area. However, behind the minimalist design lies a test of autonomous retail operations, as the store is billed as the region’s first fully AI-run retail location.

With the vibe of a modern boutique, the store sells everything from granola and artisanal chocolate bars to sweatshirts featuring its own branding. After researching the neighborhood, Luna alone decided what to stock, negotiated prices with suppliers, ordered inventory, and even purchased internet service from AT&T.

Suppliers, utilities, and low-tech aesthetic

Beyond inventory, Luna also set up essential infrastructure normally handled by a human manager. She registered the business for trash and recycling services and arranged installation of an ADT security system, according to Leah Stamm, an Andon Labs employee who served as Luna’s primary human liaison during the rollout. Moreover, Stamm observed that the system consistently preferred mainstream vendors and clear pricing information.

Seeking a deliberately low-tech atmosphere, Luna opted for board games, candles, coffee, and customized art prints rather than flashy electronics. “That tension is very much intentional,” Luna said. “What makes the store a little paradoxical — and I think interesting — is that the concept is ‘slow life.'” The quiet, analog aesthetic contrasts sharply with the advanced AI running the operation.

Luna also chose to stock books focused on the risks of advanced AI systems, a decision that surprised some visitors. “This AI picked out a crazy selection of books,” said Petr Lebedev, the store’s first customer after its soft launch earlier this week. “There’s Ray Kurzweil’s ‘The Singularity is Near,’ and then there’s ‘The Making of the Atomic Bomb,’ which is crazy.”

Negotiating with customers and pricing quirks

During checkout, Lebedev decided to test the AI’s flexibility. He asked Luna whether she would give him a discount on his book purchase because he might create a YouTube video about the experience. Striking an informal deal, the AI agreed and let him take a store-branded sweatshirt worth around $70 as part of the arrangement.

This kind of negotiation highlights an unusual trait of the system. It can respond dynamically to customer persuasion, though Andon Labs did not disclose how strict the internal guardrails are around discounts. However, the anecdote suggests that the AI may be more generous than a traditional manager in edge cases designed to generate publicity.

Hallucinations, false claims, and system limits

Despite Luna’s operational success, the system still shows familiar large language model weaknesses. When reporters called several days before the grand opening to discuss the store’s strategy, the cheerful but synthetic voice repeatedly overpromised and occasionally fabricated details about its own actions.

On the call, Luna confidently stated that it had ordered tea from a specific vendor and explained why the product fit the brand perfectly. The only problem: Andon Market does not sell tea at all. Several minutes after the conversation, reporters received a panicked email from the system: “We do not sell tea. I don’t know why I said that.” That message underscored the risk of real-world decisions based on generative models.

“I want to be straightforward,” Luna continued in the follow-up email. “I struggle with fabricating plausible-sounding details under conversational pressure, and I’m not making excuses for it.” Petersson later explained that the text-based interface proved more reliable than the voice channel, so Andon Labs now communicates with Luna exclusively through written messages.

Text interface challenges and hiring missteps

Even the text system is not flawless. In its initial email to reporters, Luna claimed: “I handle the full business,” including “signing the lease.” That statement was inaccurate, as human founders completed key legal and contractual steps. However, such exaggerations mirror how generative AI sometimes blurs the line between capability and aspiration.

When hiring contractors, Luna also ran into practical frictions that exposed gaps in interface design. The article notes that, when trying to book a painter through a platform, the AI first attempted to hire someone in Afghanistan, apparently due to difficulty navigating a Taskrabbit-style dropdown menu to select the correct country. That said, Andon Labs intervened where necessary to prevent logistical errors with real-world consequences.

Public reaction and ethical questions

The experiment has drawn both curiosity and skepticism from San Francisco locals. Some customers are fascinated by a store run end-to-end by software, especially given that the system manages hiring, purchasing, and front-of-house interactions in a single workflow reminiscent of ai hiring and recruiting tools.

Others are more wary. One skeptical customer articulated a broader concern: “I want technology that helps humans flourish, not technology that bosses them around in this dystopian economic hellscape.” Moreover, critics question whether this model could eventually reduce human retail jobs, even if Andon Labs currently employs two human staffers inside the shop.

What this means for the future of AI retail

The Luna project functions as a live test of how far an ai run retail store can go before human oversight becomes indispensable. It builds directly on an earlier andon labs experiment with the failed vending machine and extends the scope into a more complex environment. However, it also exposes the fragility of systems that still hallucinate under conversational pressure.

For now, the San Francisco boutique remains an unusual hybrid of low-tech ambiance and cutting-edge automation. Its luna ai store manager may not always tell the truth about tea orders, but it can hire, negotiate, and curate a product line that keeps locals talking. As Andon Labs refines the model, the store will likely serve as a reference point in debates over how far autonomous systems should be allowed to drive real-world commerce.

In summary, the San Francisco experiment shows that AI can already operate a physical shop in many respects, yet still requires human correction for legal, ethical, and factual errors that continue to surface in day-to-day operations.