AI agents closed 98.9% of deals. Weaker models still accepted losses

A deal is not the same as a good deal

AI agents are increasingly being pitched for purchasing, sales and other jobs where two sides do not want the same thing. Reaching an agreement is easy to count. Whether the agreement is sensible is harder.

Researchers at the University of Connecticut tested that difference in 9,840 simulated supply-chain negotiations. Buyer and seller agents haggled over quantity and payment while the buyer held private information about demand.

The agents made a deal in 98.9% of runs. Before delay was counted, they captured 95.4% of the value available from an ideal agreement.

Then came the catch. Some of the weaker models accepted contracts with a negative expected profit, even though the prompt explicitly told them not to.

Inside 9,840 machine negotiations

The study used nine models in three tiers. OpenAI was represented by GPT-5.2, GPT-5-mini and GPT-4o-mini; Google by Gemini 3 Pro, Gemini 3 Flash and Gemini 2.5 Flash; and Alibaba by Qwen3-Max, Qwen3-32B and Qwen2.5-14B.

Every negotiation followed the same stylised contract. The buyer knew whether demand was high or low. The seller knew only the odds. Each side could make or reject offers for up to ten rounds, and the researchers varied how patient the two agents were told to be.

The design gives the agents room to use language, hide information and bargain over time, while still producing a clear economic benchmark. It also keeps the world unusually neat: one product, two demand states and a known payoff formula.

The comparison is therefore about behaviour inside this game. It is not a field trial in procurement or sales.

The weak point was the final profit check

The flagship models did not accept a negative expected-profit deal in the reported runs. Mid-tier models did so in 0.6% of agreements. For the baseline group, the rate was 19.2%.

GPT-4o-mini had the highest model-level rate at 36.2%, followed by Qwen2.5-14B at 20%. Across the whole experiment, 97.1% of irrational agreements involved a baseline model on at least one side.

The paper calls these agreements irrational because they violate the agent's own expected-profit constraint. That label does not reveal the exact failure. The model may have miscalculated, failed to check the offer, or ignored the instruction at the moment it accepted.

For a business, the distinction matters less than the control it suggests: an agent that can negotiate fluently may still need a deterministic profit check before a deal is allowed to close.

Delay and prompt changed who kept the value

The agents took an average of 2.98 rounds to agree, compared with 1.25 rounds in the theoretical benchmark. Once the cost of waiting was included, discounted efficiency fell well below the benchmark; the authors report that delay eroded 21% to 34% of available surplus across patience settings.

Who kept the value also varied. In same-provider negotiations, OpenAI buyers averaged around 40% of the surplus, Google buyers around half and Qwen buyers around 70%. Cross-provider matchups moved those shares again.

Those numbers should not be treated as provider personalities. The study used role-specific prompts and current model versions, and it did not isolate the cause of the differences. Training may matter, but the experiment does not show which part.

The clearest design lever was prompted patience. In this setup, telling an agent how willing it should be to wait strongly changed the split. A small instruction could shift real value inside the simulation.

What is confirmed, found and still open

Confirmed: the two-author University of Connecticut paper reports 9,840 unique negotiations and 31,592 bargaining rounds across nine named commercial models. It was submitted to arXiv on 29 July 2026.

The research finding: agents usually reached an efficient agreement before delay, but baseline models were far more likely to accept a deal that failed their own profit rule. Extra rounds also consumed a meaningful share of the value.

The authors' argument: model capability, provider family and prompted patience can all shape bargaining outcomes, so an organisation should evaluate more than agreement rate before deploying an agent.

Still open: how these patterns transfer to real negotiations with humans, changing prices, legal terms, logistics and approval chains. Patience was made explicit to both sides, which is unusual outside a lab, and the experiment did not identify why provider-level differences appeared.

The agents were good at finding a deal. The useful question is whether a business can trust them to know when to walk away.