What Anthropic's Project Vend Teaches Us About AI in Business

The Experiment

In early 2025, Anthropic launched one of the most ambitious AI experiments in recent memory: give Claude (their AI model) actual control over a vending machine business. Real inventory. Real customers. Real money changing hands.

The results were fascinating — and messy. The AI made some brilliant calls: it identified that protein bars were underselling in gym locations because of poor positioning, it optimized pricing in hospital locations based on time-of-day demand curves, and it predicted restocking needs with surprising accuracy.

But it also made mistakes. It once ordered 200 units of a seasonal item right as demand was collapsing. It struggled with supplier negotiations. And it had no intuition for the “soft” parts of the business — relationships with location managers, handling customer complaints, reading the room.

The Three Big Lessons

1. AI is incredible at pattern recognition — if you give it data

The most impressive results came when Claude had access to historical sales data. It spotted patterns that experienced operators miss: seasonal shifts, day-of-week demand curves, cross-product correlations (“when energy drink sales spike, protein bar sales follow 24 hours later”).

The takeaway for operators: your sales data is a gold mine you're not mining. Even basic tracking — what sold, when, where — gives AI enough to make meaningful recommendations.

2. AI shouldn't replace operators — it should augment them

Project Vend tried to make AI the “operator.” That was the wrong frame. The AI wasn't good at the human parts: negotiating with location managers, handling machine repairs, deciding when to invest in new equipment.

The right model is AI as your business intelligence layer — it analyzes, recommends, and alerts, but the human makes the final calls. Think of it like GPS navigation: it tells you the optimal route, but you're still driving.

3. The tooling gap is the real opportunity

Anthropic built custom infrastructure for Project Vend. Most vending operators have... a spreadsheet. Maybe. The gap between what AI can do and what operators have access to is enormous.

That's exactly why we built VendBrain. Not to replace operators, but to give every operator with 5 machines and a pickup truck the same AI insights that Anthropic had with their multi-million dollar experiment.

What This Means for You

If you're a vending operator, the Project Vend experiment validated something important: AI can meaningfully improve vending operations. The question isn't whether AI works for vending — it's whether you have access to tools that make it practical.

The average operator is underpricing by $0.25-0.75 per item, stocking 2-3 products that should be swapped out, and making 30% more restocking trips than necessary. AI catches all of this.

You don't need a million-dollar experiment. You need the right tool.