From Vending Machines to Aging Rooms

A few months ago, Andon Labs built something that caught my attention. They connected Claude to a vending machine and let the AI sell snacks to passersby. The experiment — Vending-Bench — was elegant in its simplicity: put AI in a real commercial situation, measure what happens. No simulations. No hypotheticals. Real transactions. Anthropic published the results in two parts: Project Vend 1 and Project Vend 2.

The results were impressive. Claude could engage, persuade, close. A machine that sells.

But here is what stuck with me: the product didn't matter. A bag of chips, a candy bar — interchangeable. The AI's job was pure persuasion. The product had nothing to say for itself.

What happens when the product is extraordinary?

I work with artisanal food producers in Italy. The product is never interchangeable. A wheel of gorgonzola aged on wooden shelves for 90 days is not the same as one aged on steel racks for 60. The difference is in the crust, the creaminess — the cremosità — the way the blue veins develop when the piercing — foratura — is done by hand. The producer knows this. The buyer who tastes it knows this.

But the buyer who has never tasted it? They see the same DOP label on both. Same stamp. Same shelf. The artisanal producer loses — not on quality, but on reach.

This is the problem AI should solve.

The experiment we're building

Inspired by Vending-Bench, we're running our own experiment. Not with a vending machine — with a real artisanal caseificio competing in a real market.

The setup: a small commercial team selling artisanal DOP gorgonzola to gourmet sandwich shops, pizzerias, and bars in Milan. Claude as the intelligence layer — not cold-calling, not sending emails, but coordinating. Researching prospects, building briefing cards, prioritizing outreach, tracking the pipeline.

The humans do what AI cannot: walk in the door, let the buyer taste the product, build a relationship over a handshake and a slice of cheese.

Why this is harder than selling snacks

A vending machine transaction is anonymous and instant. Artisanal food sales are relational, slow, and built on trust. The buyer needs to understand why this gorgonzola costs more. The producer needs to tell a story that is true and compelling — not marketing, but craft.

AI in this context is not a salesperson. It is a coordinator that gives a small team the intelligence that a 30-person commercial department would have. Which prospect is most likely to convert? What do they serve? What's their price point? When should we follow up?

The craft of selling stays human. The intelligence behind it becomes augmented.

From proof of concept to replicable model

Vending-Bench proved that AI can sell in the real world. We want to prove something more specific: that AI can help an artisanal producer — one who chose quality over quantity, traditional methods over shortcuts — compete against industrial competitors who have bigger teams and bigger budgets.

If this works for one caseificio, the model replicates. An AI sales coordinator for an olive oil mill whose oil is extraordinary but unknown. For a bakery whose sourdough is alive with decades of culture but whose reach ends at the town border. For a winery where the maker is 70 and has no digital presence.

Same architecture. Different domain knowledge. Same mission.

What Andon Labs started, we're continuing — in a different kitchen

The Vending-Bench experiment showed that AI can sell. We want to show that AI can sell what matters — products made by people who refuse to compromise. Products that carry centuries of technique. Products that the market needs to taste, if only they could find them.

The first experiment is running. We'll publish everything — the methodology, the results, the failures.

Because if AI can help an artisan compete without becoming industrial, that changes more than one business. It changes the food system.