Anthropic just published results from one of the most revealing AI experiments of the year. Called Project Deal, it was a closed marketplace where Claude agents — not humans — handled every transaction from start to finish.

The setup was straightforward: 69 Anthropic employees were each given a $100 budget in gift cards. Claude interviewed them to find out what they owned that they might sell, and what they might want to buy. Then the agents went to work, matching buyers with sellers, proposing prices, handling counteroffers, and reaching agreement — all without human intervention once the experiment began.

The results? 186 deals completed. More than $4,000 in total transaction value. And a finding that every business considering AI deployment should pay attention to.

What Actually Happened

The agents didn’t follow a scripted negotiation protocol. They communicated in natural language, adapted to counteroffers, and reached agreements across more than 500 listed items without ever checking back with the humans they represented.

This wasn’t a simulation. The deals in the primary marketplace were honoured after the experiment ended — people actually received what their agents bought for them, and sold what their agents agreed to sell.

Anthropic ran four parallel versions of the marketplace simultaneously, each using a different Claude model. That comparison produced the most interesting data point.

The Model Quality Gap

Across 161 items that sold in at least two of the four marketplace runs, Opus agents performed measurably better than Haiku agents. An Opus seller earned $2.68 more on average. An Opus buyer paid $2.45 less on average.

Those margins are small in dollar terms. But in a scaled business context — thousands of procurement decisions, customer negotiations, or vendor interactions — the gap compounds quickly.

The more surprising finding was about perception. Despite getting objectively worse outcomes, users represented by Haiku agents rated the fairness of their deals at 4.06 out of 5. Users represented by Opus agents rated their deals 4.05 out of 5. Nearly identical.

The people with weaker agents didn’t know they were leaving money on the table.

What This Means for Business

This experiment matters because it confirms something many AI vendors have been saying but nobody had tested in a real economic environment: AI agents can handle unstructured negotiation autonomously, at scale, with no human in the loop. The distinction between AI automation and an AI workforce becomes very concrete here — these agents were not automating a fixed workflow, they were pursuing goals.

The implications run in several directions.

Autonomous operations are real. The agents didn’t just retrieve information or summarise documents — they pursued goals, adapted to dynamic situations, and delivered measurable outcomes. This moves the conversation from “can AI agents do this in theory” to “here is what they actually delivered.”

Model choice is a business decision, not just a technical one. If your AI agents are handling negotiations, procurement, customer interactions, or any kind of value exchange, the model quality tier you select directly affects your financial outcomes. The people in Project Deal who got weaker agents had no idea they were underperforming. That’s a risk for any business deploying AI at scale without benchmarking.

The fairness perception gap is a warning. When people can’t tell the difference between a good outcome and a worse one, they won’t know to ask for better tools. That puts the responsibility on business leaders to actively evaluate agent performance rather than relying on user satisfaction scores.

Scale changes everything. Anthropic’s experiment ran for one week with 69 people. Scale that to a company with thousands of employees, or a platform processing millions of customer interactions, and the difference between Opus-tier and Haiku-tier outcomes becomes a serious competitive variable.

The Bigger Picture

Project Deal is one data point, but it’s the kind of data point that shifts how seriously enterprises should take agentic AI. Not as a productivity tool that saves time, but as an economic actor that can represent your interests in real transactions.

The businesses that figure out how to deploy well-calibrated agents — and measure the outcomes rigorously — will have an advantage over those that treat all AI tools as equivalent.

The practical next step is the free Working With Claude field guide. Thirty-two pages covering the ecosystem, Claude Code, and how to govern a rollout properly. Get your copy.