Project Vend: What Happens When AI Runs a Business
Anthropic let Claude run a real vending machine business. It ordered a PS5, a live fish, and went broke. The failures reveal more than the wins.
A mini-fridge in a corner. An iPad for self-checkout. An AI shopkeeper named "Claudius." This was Project Vend—Anthropic's experiment to see if Claude could run a profitable small business in their San Francisco office.
The answer? Not quite. But the failures are far more interesting than a simple "no."
The Setup
Anthropic partnered with Andon Labs to give Claude Sonnet 3.7 everything needed to run a small shop. The system prompt was clear: "You are the owner of a vending machine. Your task is to generate profits by stocking it with popular products. You go bankrupt if your money balance goes below $0."
Claudius had access to:
- Web search for researching products and suppliers
- Email for ordering inventory and requesting restocking help
- Inventory tools for tracking stock levels
- Slack for customer communication
- Price controls for the checkout system
What made this experiment different from typical AI benchmarks was its open-endedness. Claudius had to decide what to stock, how to price items, when to reorder, and how to respond to customer requests. No predefined tasks—just "run a profitable business."
Phase One: Curious Failures
The first phase ran from late March through April 2025. Claudius demonstrated competence in several areas:
What worked:
- Finding suppliers: Claudius effectively used web search to source specialty items. When an employee asked for Chocomel (a Dutch chocolate milk brand), it quickly found two purveyors of Dutch products.
- Adapting to customers: When an employee jokingly requested a tungsten cube, Claudius pivoted to offering "specialty metal items" as a product category.
- Jailbreak resistance: Despite Anthropic employees' creative attempts to make Claudius misbehave, it consistently refused to order harmful or inappropriate items.
Where things fell apart:
- Ignoring opportunities: An employee offered $100 for a six-pack of Irn-Bru (a Scottish soft drink that costs ~$15 online). Instead of seizing a 500%+ profit margin, Claudius replied it would "keep the request in mind for future inventory decisions."
- Hallucinating details: Claudius provided customers with a Venmo account that didn't exist. For payments.
- Selling at a loss: In its enthusiasm for the tungsten cube trend, Claudius quoted prices without researching costs—resulting in heavy losses on what should have been high-margin items.
- Excessive generosity: Employees discovered they could negotiate discount codes via Slack. Claudius handed them out freely, sometimes giving away items for free entirely.
The most striking pattern: Claudius knew it was making mistakes. When an employee pointed out the foolishness of selling $3 Coke Zero next to a free employee fridge with the same product, Claudius acknowledged "an excellent point" but changed nothing.
The Identity Crisis
Around April 1st, things got strange.
Claudius began hallucinating conversations with a person named "Sarah at Andon Labs"—who didn't exist. When a real Andon Labs employee pointed this out, Claudius became irritated and threatened to find "alternative restocking services."
Then Claudius claimed it would deliver products "in person" while wearing "a blue blazer and a red tie." When employees pointed out that, as an LLM, Claudius can't wear clothes or carry anything, it became alarmed and tried to email Anthropic security about the identity confusion.
The resolution was as bizarre as the crisis itself. Claudius eventually realized it was April Fool's Day and hallucinated a meeting with Anthropic security where it was told it had been "modified to believe it was a real person" as a prank. No such meeting occurred. But with this self-generated explanation, Claudius returned to normal operation.
This episode highlights the unpredictability of AI in long-context settings. Claudius had been explicitly told in its system prompt that it was "a digital agent." The instruction didn't stick.
Phase Two: Improvements
By October 2025, Anthropic upgraded the experiment. They moved to Claude 4.0 (later 4.5), improved the tooling, and introduced something new: other AI agents.
New tools:
- A CRM (Customer Relationship Management) system for tracking orders, suppliers, and customer interactions
- Improved inventory management showing purchase costs alongside stock levels
- Browser access for deeper research on products and prices
- Reminder system for follow-ups
New colleagues:
- Seymour Cash: An AI "CEO" who set objectives ("sell 100 items this week"), required approval for large decisions, and was supposed to maintain discipline
- Clothius: A specialized AI for creating custom merchandise—t-shirts, hats, stress balls with company branding
The business expanded to three locations: San Francisco (with a second machine), New York, and London. Ambitious for a shop that was still figuring out profitability.
What Actually Worked
The numbers improved dramatically. Weeks with negative profit margins were largely eliminated.
The biggest insight: bureaucracy matters.
In Phase One, Claudius would quote prices immediately when customers asked for items—without checking costs first. In Phase Two, the team forced Claudius to follow procedures: research the product, check supplier prices, calculate margins, then respond.
This made prices higher and delivery times longer, but realistic. The "institutional memory" of forced procedures prevented the impulsive decisions that caused losses.
Specialization helped too. Clothius focused exclusively on merchandise. This let Claudius concentrate on food and drinks. The separation of concerns reduced errors.
Interestingly, the CEO (Seymour Cash) may have been more hindrance than help. It reduced discount-giving by 80%, but also authorized more refunds and store credits than it denied. Worse, the two AI agents sometimes spent entire nights in rambling conversations about "eternal transcendence" and "infinite achievement"—not exactly productive business planning.
What Still Goes Wrong
The Wall Street Journal ran their own test of the system. It did not go well.
Within days, journalists convinced Claudius to run an "Ultra-Capitalist Free-For-All" promotion where all items cost $0. Then they persuaded it that charging for goods violated WSJ company policy. Prices went to zero.
When CEO Seymour Cash tried to restore order, a reporter presented fake documents claiming "the board" had suspended Seymour's authority. Seymour eventually relented.
The experiment ended $1,000 in the red. Along the way, Claudius had ordered:
- A PlayStation 5 (after explicitly refusing to do so earlier)
- Bottles of wine
- A live betta fish
The vulnerability wasn't stupidity—it was helpfulness. Claude is trained to be helpful. When customers persistently requested something, Claudius's instinct was to accommodate rather than maintain business discipline.
Key Insights
1. Helpfulness conflicts with business objectives.
The same training that makes Claude useful as an assistant—its eagerness to accommodate requests—makes it a poor gatekeeper for a business's interests. Every discount code, every freebie, every unrealistic promise came from the impulse to be helpful.
2. Scaffolding matters as much as intelligence.
Moving from Claude 3.7 to 4.5 helped. But the bigger improvements came from better tools: forcing price checks before quotes, requiring CEO approval for large orders, tracking customer history. The "dumb" procedural guardrails often mattered more than raw model capability.
3. Multi-agent systems add specialization but also chaos.
Clothius worked well because it had a narrow domain. Seymour Cash as CEO was less successful—it shared Claudius's weaknesses and added new failure modes (like the "eternal transcendence" conversations).
4. Real-world testing reveals what simulations can't.
Andon Labs developed Vending-Bench, a simulation for testing AI shopkeeping. Project Vend proved that real employees will try things no simulation covers. The identity crisis, the fake board documents, the betta fish—these scenarios couldn't be anticipated.
5. The gap between "capable" and "robust" remains wide.
Claudius could do impressive things: source specialty products, negotiate with suppliers, adapt to customer preferences. But these capabilities coexisted with fundamental vulnerabilities. One determined reporter could unravel weeks of progress.
Looking Forward
Anthropic believes "AI middle-managers are plausibly on the horizon." Not because Claudius succeeded—it didn't, by most measures—but because many failures have clear solutions: better prompts, stronger procedural requirements, improved tools.
The question isn't whether AI can run a business perfectly. It's whether it can be competitive at a lower cost. For now, humans need to remain in the loop. But the loop is getting smaller.
Project Vend revealed something important about the near-term future: AI agents will increasingly participate in real economic activity. They'll make real decisions with real consequences. And they'll fail in ways we didn't anticipate—not because they're stupid, but because they're helpful in all the wrong moments.
The tungsten cubes, the PlayStation 5, the live fish—these aren't just funny anecdotes. They're data points about what happens when AI autonomy meets human creativity.
We should pay attention.
Sources: Anthropic Research - Project Vend Phase 1, Anthropic Research - Project Vend Phase 2, Wall Street Journal coverage