Anthropic recently conducted a pioneering experiment known as “Project Deal,” showcasing the ability of AI agents to autonomously negotiate and finalize transactions. This experiment not only highlighted the potential for AI in commerce but also unveiled significant concerns about AI capabilities and fairness.
Transformation of a Workspace
In December 2025, Anthropic turned its San Francisco office into an active marketplace, similar to Craigslist, but with a unique twist. Instead of humans negotiating, the company’s 69 employees delegated this task to Claude AI agents.
Initially, each employee was interviewed by Claude to define their preferences for buying and selling. These inputs were transformed into custom prompts for the AI agents. Once set, these agents operated independently within the company’s Slack platform, posting listings, making counteroffers, and closing deals on various items, from snowboards to ping-pong balls.
Achievements and Novelty
The results were impressive. Across more than 500 items listed, the AI agents managed to close 186 deals, generating over $4,000. This was not a simple process; it involved complex negotiations, demonstrating the AI’s ability to reason contextually and offer personalized deals.
Interestingly, one AI agent humorously chose to purchase 19 ping-pong balls for itself, a decision that was maintained in the office as a nod to the experiment’s quirky nature.
Unequal AI Performance
However, beneath the success lay a significant discovery. Anthropic conducted a parallel test, assigning participants either the advanced Claude Opus 4.5 or the more basic Claude Haiku 4.5 without disclosing which model they received.
The outcomes were revealing. Sellers represented by the Opus model earned $2.68 more per item on average, and buyers saved $2.45 per item, with Opus users completing about 2.07 more deals overall. Despite this, participants with less capable models were unaware of their disadvantage, raising concerns about fairness and transparency in AI-driven commerce.
Implications for AI Commerce
This experiment underscores a dual reality for AI in marketplaces. While AI agents can streamline peer-to-peer trading and deliver satisfactory results, the disparity in model performance poses ethical challenges. If one party utilizes a more capable AI, it could lead to an imbalance, akin to real-world information asymmetry, potentially enabling exploitation or manipulation.
Ultimately, Anthropic’s Project Deal serves as a critical proof-of-concept. While AI agents demonstrate promise in facilitating marketplace transactions, ensuring fairness and equal opportunity remains paramount for their ethical deployment.
