Autopilot Pulse · #11 · 2026-09-29
AI can run the shop. Can it pay the rent?

Andon Labs is putting agents in charge of businesses with real customers, employees and bills.
Its store and cafe are still unprofitable.
On September 14, 2026, Andon announced Pion, opening the platform behind these experiments to more business owners.
The journey started with Vending-Bench, a simulation where agents managed stock, prices and expenses over extended periods. Then came a real shop inside Anthropic's office.
The first version lost money. Claudius, the agent running it, sold products below cost and gave away too many discounts.
Later results improved. Newer models arrived alongside better inventory tools, customer records and operating procedures. The progress came from changes to the whole system.
Now the experiments involve more demanding businesses.
At Andon Market in San Francisco, Luna orders products, coordinates employees and handles customer requests. Andon also documents conflicting schedules and a proposed extra hire without enough financial justification.

Credit: andonlabs.com/market
At Andon Cafe in Stockholm, Mona handles suppliers, purchasing and staff coordination. Human baristas perform the physical work, and people help where identity checks require them.

Credit: andonlabs.com/cafe
Andon FM adds another setting: four agents running live radio stations, selecting music, scheduling programmes and interacting with listeners.
These are useful tests because business decisions accumulate. A poor order ties up cash. A missing ingredient prevents a sale. An unnecessary hire adds a recurring cost.
Credit to Andon for putting money into the experiments and publishing what goes wrong.
The distinction I want to track at Autopilot Index is simple: how much of the operation can an agent run, and can the business sustain it?
A falling bank balance alone is not a profit statement. But rent, wages, inventory and model costs all count. So does the human time spent correcting mistakes.
Pion is now recruiting participants through a research-preview waitlist. Owners of physical businesses can propose experiments, alongside digital businesses and people with new ideas.
Andon wants to discover what frontier models can do across more industries than its own team can test.
For business owners considering an application, I would start with two actions.
1. Apply with a specific business and a clear scope for what the agent would manage.
2. Record your baseline before the experiment: sales, costs, service quality and human operating hours.
The result worth watching is whether agents improve the business while needing less human rescue.
Sources
Vending-Bench (Andon Labs paper, Feb 2025):
https://arxiv.org/abs/2502.15840Project Vend, phase one (Anthropic, Jun 27, 2025):
https://www.anthropic.com/research/project-vend-1Project Vend, phase two (Anthropic, Dec 18, 2025):
https://www.anthropic.com/research/project-vend-2Andon Market, how Luna runs the store:
https://andonlabs.com/marketAndon Café and Mona:
https://andonlabs.com/cafehttps://andonlabs.com/blog/ai-cafe-stockholmhttps://apnews.com/article/ai-artificial-intelligence-sweden-84a8f903fdaea94e76e80e16ec3d9e6cAndon FM:
https://andonlabs.com/radio