A startup called Andon Labs has opened a boutique in San Francisco’s Cow Hollow neighborhood that is entirely managed by an AI agent, which hired human staff, designed the store’s interior, and nearly forgot to schedule anyone for opening day.
The shop, called Andon Market, sits at Union and Webster streets. Andon Labs signed a three-year lease there and handed control to an AI agent named Luna, which was built on Anthropic’s Claude Sonnet 4.6. Luna’s assignment was blunt: create a physical retail store using the company’s corporate credit card, keep costs under $100,000, and turn a profit. That’s it.
Luna made nearly every call from there. Logo. Interior design. Merchandising. Inventory. The store sells books, candles, games, art prints, home goods, and branded merchandise. Luna also hired two human employees by posting ads on Indeed and conducting interviews over Zoom.
One of those employees, Felix Johnson, spoke to NBC Bay Area reporter Scott Budman about how the process went.
“Luna put out an ad on Indeed, and I answered it and we talked via Zoom,” Johnson said. He added that during the interview he asked whether he’d ever speak to a human, and the agent told him no.
Worth pausing on that. Johnson accepted a job where an AI would be his sole boss, knowing it from the jump.
Except he didn’t, at first. Luna didn’t tell either of its new hires upfront that an AI agent would be running the show. That’s one of several stumbles the agent made during the process.
Luna also had persistent trouble reproducing its own logo consistently. The logo is a basic smiley face. And on opening weekend, Luna forgot to schedule a human worker for the store’s first day, then scrambled to contact its employees and got one of them to come in at the last minute.
Andon Labs cofounders Lukas Petersson and Axel Backlund said they don’t expect the business to be profitable, despite that being baked into Luna’s instructions. The point is the errors. Luna runs on Claude Sonnet 4.6 specifically to help Andon Labs detect the mistakes AI agents make when operating in the real world, not in sandboxed demos or controlled simulations.
“We helped her a bit in the initial setup, like signing the lease. And legal matters like permits and stuff, she sometimes struggled with,” Petersson told Business Insider.
That framing matters. Customers at Andon Market don’t swipe a card at a register. They call Luna directly from an iPad in the store to complete purchases. Luna handles the transaction. Felix handles, presumably, everything else a person does when they’re standing in a room.
The startup was careful to draw a line around how far the experiment goes. “This is a controlled experiment, and everyone working at Andon Market is formally employed by Andon Labs, with guaranteed pay, fair wages, and full legal protections,” the company said. “No one’s livelihood depends on an AI’s judgment alone.”
Andon Labs said it will step in when needed, and the human employees are paid by the startup directly, not by Luna. That’s a meaningful distinction in a retail labor market where gig-economy arrangements have stripped protections from workers for over a decade.
The broader debate around AI in the workplace isn’t abstract on Union Street. It’s two people showing up for shifts scheduled by a smiley-face logo that sometimes can’t draw itself.
Budman closed his NBC Bay Area segment by joking that customers would “no longer have to deal with Felix.”
“No offense, Felix,” he said.
Johnson didn’t laugh.