Anthropic conducted an internal experiment, Project Deal, to test how AI agents perform as buyers and sellers in a controlled marketplace. The setup simulated real transactions while allowing the company to observe how different AI models behave in economic interactions.

Experiment Structure And Notable Outcomes

The pilot involved 69 employees, each given a $100 budget in gift cards. Over the course of the experiment, the system facilitated 186 transactions with a total value exceeding $4,000. Anthropic tested four marketplace formats, including one “real” environment where participants were fully represented by its most advanced AI model and all transactions were completed after the experiment. The remaining three setups were used for research purposes to compare outcomes across different conditions.

Follow THE FUTURE on LinkedIn, Facebook, Instagram, X and Telegram

Agent Quality And Market Efficiency

Results indicate that the quality of the AI agent had a measurable impact on outcomes. Participants represented by more advanced models achieved better results in terms of pricing and deal completion. At the same time, users were generally unaware of differences between agent types, suggesting that performance gaps may not be immediately visible to participants.

Business Implications And Future Prospects

The experiment highlights the role of model capability in shaping transaction outcomes. Initial instructions given to agents had a limited impact on final prices or deal success rates. Instead, underlying model performance appeared to be the primary factor influencing results. This raises questions about how AI-driven marketplaces may ensure balanced outcomes between participants.

Charting A New Course For AI In Commerce

Project Deal provides an early view of how AI systems could operate in transactional environments. Findings suggest that agent-based commerce is technically viable, while also pointing to the need for clearer standards around model performance and transparency as adoption expands.