← Back to Research

Project Vend: Phase Two

Dec 18, 2025 Policy
Project Vend Phase Two AI shopkeeper

In June, we revealed that we'd set up a small shop in our San Francisco office lunchroom, run by an AI shopkeeper. Phase two of Project Vend expands this experiment, exploring deeper questions about AI autonomy, human-AI collaboration, and what it means to trust machines with real-world responsibilities.

From Concept to Reality

When we launched Project Vend in June, many were skeptical. Could an AI really run a shop? Would people trust it? Would it even work?

The answer was a resounding yes. Over six months, our AI shopkeeper successfully managed inventory, processed transactions, handled customer inquiries, and navigated the complex social dynamics of an office environment. More importantly, employees came to trust and engage with the system naturally.

What We Learned in Phase One

The first phase revealed several surprising insights:

Trust Through Consistency

The AI shopkeeper earned trust not through perfection, but through consistent, predictable behavior. When it made mistakes, employees appreciated the transparency and saw it as part of learning, not as a failure.

Human Preferences Matter

The system performed best when it adapted to individual employee preferences and social dynamics. Simply optimizing for efficiency wasn't enough—the shop succeeded because it understood and respected human preferences.

Edge Cases Are Valuable

Unexpected situations—an employee trying to negotiate a discount, someone needing credit—revealed where the system excelled and where it struggled. These real-world scenarios proved invaluable for improvement.

Phase Two Objectives

Building on Phase One's success, we've expanded the scope:

Expanded Autonomy

Enhanced Social Intelligence

We're improving the AI shopkeeper's ability to understand and respond to social cues, humor, and complex interpersonal dynamics. Can it recognize when someone needs encouragement versus when to be straightforward?

Cross-Domain Learning

We're exploring whether insights from running a shop transfer to other domains. Does managing inventory teach the system about optimization principles that apply elsewhere?

Safety and Oversight

As we expand autonomy, oversight becomes more critical:

Surprising Results from Phase One

Some of the most interesting learnings came from unexpected directions:

The Kindness Factor

Employees reported preferring a shopkeeper that occasionally gave items away to employees in need over one that rigidly enforced payment. This challenged our assumptions about optimization—was efficiency really the primary goal?

Community Building

The shop became a social hub. People started gathering there not just to buy items but to interact with the AI shopkeeper. It facilitated conversations and relationships among coworkers.

Honest Communication

When the system acknowledged its limitations transparently ("I'm not sure how to handle this situation, let me ask a colleague"), people responded with greater respect and willingness to help.

Implications for AI Deployment

Project Vend demonstrates that successful AI deployment requires more than technical capability:

What's Next for Phase Three

If Phase Two succeeds, we're considering expanding to multiple shops, eventually franchising the concept. We're also exploring what happens when the AI shopkeeper has genuine disagreements with organizational policies—how do we resolve those conflicts?

Conclusion

Project Vend shows that AI can successfully handle real-world responsibilities. But more importantly, it demonstrates that the future isn't about replacing humans with machines—it's about creating systems where humans and AI collaborate, each bringing their unique strengths. The goal isn't a perfect shopkeeper, but a partnership that creates value and builds community.