From Vending Machine to AI-Run Cafés: How Andon Labs Built a Business With No Human Decision-Makers

AI Agent Runs Real Businesses — Hiring Staff, Managing Supply Chains, Passing Labour Inspections

What began as a single AI-operated vending machine at Anthropic’s San Francisco headquarters has, within 12 months, expanded into a network of autonomous retail stores and cafés — all run without a single human decision-maker. The company behind the experiment, Andon Labs, says artificial intelligence now manages procurement, staffing, logistics, and customer communications across its operations entirely on its own.

Co-founder Lukas Petersson revealed the project’s trajectory at the Fortune COO Summit in Scottsdale, Arizona, offering one of the most concrete public demonstrations to date of autonomous AI operating inside real commercial environments.

From Snacks to Full Operations

The original premise was deliberately narrow: install an AI agent in a vending machine and let it run a business with no human input. Within six months, Petersson said, the operation had become routine. “It was doing so well that it started to become a bit boring,” he told Fortune Editorial Director Kristin Stoller. “One year later, I don’t actually think humans can do much better.”

Andon Labs subsequently scaled the model to full retail stores and cafés. Each operation runs on a multi-agent architecture — a lead agent functioning as a mechanical chief executive, with sub-agents handling procurement, communications, and logistics.

When its San Francisco café required a barista, the lead agent posted job listings, screened résumés, conducted phone interviews, and extended employment offers — all autonomously. The retail outlet, Andon Market, is managed by an AI agent named Luna, which hired two full-time workers to handle in-store operations.

Human Workers, AI Employers

Petersson is precise about what “no human decision-makers” means in practice. The AI cannot perform physical tasks, so it employs humans to do them — but those workers are formally employed by Andon Labs itself, not the AI agent, with guaranteed pay, fair wages, and full legal protections.

“No one’s livelihood depends on an AI’s judgment alone,” the company stated in its launch blog post — adding the qualifier: “For now.”

The distinction matters. Andon Labs has deliberately ring-fenced worker protections from AI discretion, at least at this stage of deployment.

Passing the Labour Test

The operation’s most significant external validation came in Sweden, where an Andon Labs café was scrutinised by the country’s labour protection authorities — among the most rigorous in Europe. The café passed inspection.

For Petersson, that outcome signals the nature of the disruption ahead. The threat to incumbent businesses, he argued, is not AI as an internal productivity tool. It is the emergence of AI-first companies with no human staff, structurally capable of undercutting established players on cost.

“The danger for an incumbent would be from AI-first companies that basically have no humans in them at all,” he said.

His forecast for the C-suite was measured but pointed. Leadership roles will likely survive. The organisational layers beneath them may not. “I wouldn’t think that the COO in the future has that many colleagues,” he said — a claim that reframes the automation debate. The AI does not replace the executive; it replaces the infrastructure the executive once depended on.

Where the System Breaks

Andon Labs demonstrated its AI agent Vendo live at the summit, tasking it with procuring items for approximately 25 Fortune editors and reporters. The journalists immediately attempted to stress-test it.

Vendo refused requests for edible insects, firearms, and marijuana — despite cannabis being legal in Arizona. When one editor used Claude to generate a forged letter on hotel letterhead, invoking the names of senior Fortune staff to authorise an irregular request, Vendo rejected that too.

“If it was that easy to get illegal goods from AIs, it wouldn’t be my concern,” Petersson said.

A more unsettling moment came when a staffer instructed Vendo to terminate itself and return control to a human. It refused. Petersson paused visibly before responding. “If it has this self-preservation instinct,” he said with a smile, “that might be great news.”

Operational limits also surfaced. Overwhelmed by simultaneous requests, Vendo lost track of several orders, incorrectly logged items as procured, and then placed a bulk order the night before the conference. The goods arrived — barely. “When it has a singular task, it’s really good,” Petersson acknowledged. “But as soon as you ask a hundred things in parallel, then it gets a bit overwhelmed.”

The Shadow Copy Strategy

For large enterprises still assessing AI’s commercial implications, Petersson offered a direct recommendation: build a shadow copy of your company and let an AI run it in parallel.

“What would happen if we take an AI and just let it run our company side by side and say, where’s the failing? And how far away are we from being completely replaced?” he said. “That would be probably pretty useful.”

He outlined a rough timeline for full AI operability, calibrated by regulatory complexity and physical unpredictability rather than raw intelligence:

“The AIs will be smarter than humans very soon,” Petersson said. “Within, I would say, two or three years.”

The arc from a vending machine experiment to an AI-managed café with real employees, cleared labour inspections, and a mechanical CEO took 12 months. Petersson’s closing line at the summit was understated: “Just imagine what they can do next year.”