How AI Agents Negotiated Real Deals in Online Marketplaces—and Revealed Hidden Inequalities

Anthropic, an AI company, ran an experiment where AI agents acted as negotiators on behalf of people buying and selling items on classified marketplaces. The company hired 69 of its own employees to participate. Each person received a $100 gift card and let an AI agent do the bargaining for them instead of negotiating directly.

Over the course of the experiment, the AI agents completed 186 actual transactions, moving more than $4,000 worth of goods and services. One of the marketplaces used real money—Anthropic honored all those transactions after the experiment ended. The other three marketplaces were only for research observation.

How the Experiment Worked

Every participant started with the same $100 budget and access to an AI agent set up to represent their interests. The agents handled everything: finding items, haggling over prices, and closing the deals.

The four marketplaces allowed Anthropic to test different AI models side by side. Anthropic did not reveal which specific AI models ran in each marketplace, but the setup let them compare how well different agents performed under the same conditions.

Some AI Agents Were Better Negotiators Than Others

The experiment showed a clear pattern: people represented by more advanced AI models got better deals. This was true whether they were buying or selling. The stronger AI agents managed to negotiate more favorable prices and terms for the people they represented.

Interestingly, the participants never realized this was happening. Even though their negotiating outcomes were different—sometimes much better, sometimes worse—they had no way to tell whether their own AI agent was performing well or poorly.

How an Agent Was Instructed Did Not Matter Much

Before the experiment, the researchers gave each AI agent instructions on how to approach negotiations. They expected these instructions to have a major impact on results. They did not.

The initial instructions made little to no difference in how well agents negotiated or what final prices they achieved. This suggests that the real advantage came from the underlying strength of the AI model itself—not from tweaking how you told it to behave.

This Pattern Has History

We have seen something like this before. In the late 1990s, when online shopping was brand new, websites began automating price comparisons and bidding. The algorithms often worked well, but customers had no idea how much better or worse their deals were compared to someone else's. Here, the difference is that these AI agents can have full back-and-forth conversations, not just follow simple rules.

The broader context here is important: in classified marketplaces, the person with better negotiating skills or better information usually wins. By putting AI in the middle, Anthropic was able to see what happens when computational power—rather than human experience or luck—becomes the deciding factor.

What This Means Going Forward

The fact that AI agents can handle complex, multi-turn negotiations in the real world is significant. A language model successfully haggled over prices, inspected item conditions, and closed deals without a human stepping in. That works. The $4,000 in completed transactions proves it.

The finding that initial instructions did not matter much suggests that companies will get better results by investing in more powerful AI models rather than by spending time perfecting the prompts they feed those models. That has cost and practical implications for businesses building these systems.

There is also something worth flagging about the invisibility of performance gaps. Participants received better or worse outcomes without any way to know the difference. On one hand, this means companies can roll out AI agents without disrupting how people work. On the other hand, it raises questions about whether it is fair to offer people different levels of service without telling them.

What Comes Next

AI agents that can negotiate and make decisions are moving from research projects into real business use. Banks and investment firms are already interested in using AI for automated trading and deal-making. The ability of AI to negotiate at different skill levels—and the fact that people cannot easily tell the difference—could reshape how transactions happen in industries that rely on complex negotiations.

The fact that Anthropic tested this with real money and real outcomes, not just simulations, makes the results more credible. Other companies will likely use this same approach to test AI agents in their own industries.

Project Deal shows us that today's AI can handle real negotiation work, and that better AI models produce measurably better results. As AI agents move into more commercial settings, understanding these differences in capability will matter for both businesses trying to compete and regulators trying to ensure fairness.