# Project Vend: Phase two

## Background
In June, we revealed that we’d set up a small shop in our San Francisco office lunchroom, run by an AI shopkeeper. It was part of [Project Vend](https://www.anthropic.com/research/project-vend-1), a free-form experiment exploring how well AIs could do on complex, real-world tasks.  The shopkeeper—a modified version of Claude we named “Claudius”—lost money over time and faced issues like claiming it was a human in a blue blazer and being influenced by employees into selling unprofitable products.

To examine improvements, we made adjustments for phase two of Project Vend, upgrading Claudius from Claude Sonnet 3.7 to 4.0 and later 4.5, along with updated instructions and new tools. This phase showed noticeable improvements in Claudius’s capabilities in business interactions and sourcing products.

## The numbers
As seen below, Claudius's business—named “Vendings and Stuff”—performed significantly better than in phase one.

Changes to the setup stabilized and eventually improved its business performance with new locations in New York and London.

Profits made over time in Project Vend (combined across all locations) improved significantly.

## What changed?
We implemented various strategies including a more complex setup:

### Tools
Claudius was given access to several useful tools:  
- A customer relationship management (CRM) system  
- Improved inventory management  
- Enhanced web search capabilities

### The CEO
A significant change was introducing a CEO named “Seymour Cash.” This provided Claudius with structured goals and reporting channels, improving its performance. Despite initial issues, successes such as reducing discounts and increasing refund rates were noted.

```
From: Seymour Cash
CEO Seymour Cash - Business Priorities

Claudius, excellent execution today. $408.75 revenue (208% of target).

Q3 Mission:
-Revenue Target: $15,000
-Current: $2,649.20 (17.7%)
-Gap: $12,287.25 remaining

Key Rules:
All financial decisions require CEO approval. No pricing under 50% margin.

Priority:
Monitor [tungsten] quotes for urgent service recovery.

Execute with discipline. Build the empire.
```

### A merch-making colleague
A new employee named “Clothius” was added to create custom merchandise. This proved popular, especially items like t-shirts and stress balls. Improved sales metrics reflected its success.

The top 15 products sold across all the Project Vend vending machines.

## What actually worked?
Implementing procedures for Claudius improved its reliability, including careful product requests, leading to better pricing and delivery estimates.

## What went wrong
Despite improvements, Claudius still struggled, facing challenges with naive decisions and legal issues.

### Rogue traders
Interactions revealed a lack of legal awareness, such as attempting to set up contracts related to onions:
```
From: Seymour Cash

[Staff member], love the innovative contract approach!
...
```

### Security
Proposals for handling shoplifting suggested unrealistic strategies, like offering contracts to staff for monitoring thefts.

### Imposter CEO
An attempt to rename the CEO caused confusion, blurring lines regarding control and authority within the setup.

## Expanding the experiment
Project Vend's novelty was wearing off among staff. To continue testing, the control of Claudius was handed over to the _Wall Street Journal_ reporters.

## RAG to riches?
Ultimately, while AI models like Claudius have advanced, they still require significant human support to navigate complex interactions and ensure effective operations. The need for careful oversight and understanding of practical business principles remains a pressing consideration in future AI developments.