AI & Models
Anthropic tests AI-to-AI commerce in internal pilot
Anthropic’s Project Deal pilot found that while advanced AI agents secure objectively better outcomes in commerce, users may not perceive the performance disparity.
Anthropic has conducted an internal pilot experiment called Project Deal, establishing a classified marketplace designed for agent-on-agent commerce—a setup where AI agents acted as both buyers and sellers. The test allowed these agents to represent participants to negotiate and strike real deals for real goods and real money.
The pilot experiment operated under the following parameters:
- Participants: 69 Anthropic employees made up a self-selected participant pool.
- Budget: Each employee was given a budget of $100, which was paid out via gift cards, to purchase items from their co-workers.
- Transactions: The pilot resulted in 186 deals.
- Volume: The total value of the completed transactions reached more than $4,000.
To study the dynamics of these negotiations, Anthropic ran four separate marketplaces utilizing different models. One of these setups was real, meaning every participant was represented by the company’s more-advanced model, and the negotiated deals were actually honored after the experiment concluded. The other three marketplaces were conducted for study purposes.
According to Anthropic, the experiment revealed that users represented by more-advanced models achieved superior results in their negotiations. However, the human users did not seem to notice this performance disparity. This finding raises concerns about the emergence of agent quality gaps. As Anthropic, an AI research company, noted, “people on the losing end might not realize they’re worse off.”
Anthropic characterized the initiative as a pilot experiment with a self-selected participant pool. Despite the limited scope, the company stated it was surprised by how well the project functioned. Interestingly, the initial instructions given to the AI agents didn’t appear to affect either the likelihood of a sale or the final negotiated prices.
Why it matters
Anthropic’s experiment highlights a critical blind spot in AI-driven commerce: as agents negotiate on our behalf, performance disparities may create agent quality gaps that users are unable to detect.