top of page

AI Sale Negotiations

Nov 3
4 min read

Our research team conducted a study on how AI perform sales negotiation. The paper was published at CHI 2026. Link: https://arxiv.org/abs/2603.22766


A negotiation issue means one item to agree on, like rent. Many tools ask people to track several issues at once. In procurement, that could be price, delivery, warranty, and payment terms. In contracts, it could be scope, renewal, liability, and timing.


Our question was simple. What happens when the number of issues grows?


Eye-level view of handwritten option cards on a kitchen table beside a small speaker
A simple way to think about many negotiation issues.

What we did



We asked 32 people to play a tenant. They negotiated a rental with a GPT-4 landlord. We changed how many negotiation issues were in the deal. Each person completed all conditions.


We also tested a Decision Support tool.


The tool had two parts.


The first part was a grid. It showed which options the AI landlord was likely to accept. To build this grid, we used Bayesian updating. Bayesian means updating a guess each time new evidence arrives. After each offer, the tool updated its view of what the AI might accept.


The second part was a progress bar. A zone of possible agreement means the options both sides could accept. The bar showed how narrow that zone was and who it favoured.


The tool did not tell the tenant what to offer. It only showed what was still possible.


Here is the short version of the study design.


What we changed


  • The number of negotiation issues 1, 3, 5, or 7

  • Whether the person had the Decision Support tool With or without the tool


What we measured


  • Payoff

  • Distance to the best deal

  • Mental workload


A best possible joint deal means a deal where neither side can gain without the other losing. We looked at how close each outcome came to that kind of deal.


What we found without support


Without help, tenant payoff did not drop right away.


It stayed fairly stable from 1 to 3 issues. But the change from 1 to 3 was close to significant, with p = .054 and a medium effect. So the early part was not perfectly flat.


Then the pattern changed. Payoff dropped sharply at 5 and 7 issues.


A plateau-cliff effect means stable at first, then a sudden drop. That was the pattern we saw in the no-tool condition.


This matters for tool design because many interfaces still treat extra deal terms as a small change. Our result suggests that adding issues can change the task itself. At some point, the user is no longer just choosing between options. They are trying to remember trade-offs, infer what the AI will accept, and avoid giving up value by mistake.


What changed with the tool


With the Decision Support tool, payoff did not fall significantly as issues increased.


The biggest gains came when the task had more issues. At 7 issues, deals were 12.4 payoff units closer to the best possible joint deal. At 5 issues, they were 8.94 closer. At 3 issues, they were 5.50 closer. At 1 issue, the difference was 0.09.


That pattern fits the role of the tool. When there is only one issue, there is not much to track. When there are more issues, the grid and progress bar help people see what remains possible.


Joint payoff was higher with the tool. People also used fewer turns and gave faster first replies. They reported lower mental demand and effort. They rated usability 81.02 on a 100 point scale.


The tool did not replace judgement. It did not say, “Make this offer.” It showed the live shape of the deal space.


Close-up view of colored tiles arranged as possible deal options on a wooden board
The tool showed possible options without choosing for the user.

What this means for AI deal tools


For product managers, designers, and AI engineers, the useful lesson is not that every negotiation needs an assistant.


The lesson is narrower. In this rental task, support became more useful as people had to track more issues. The tool helped by making the remaining deal space visible.


That is different from giving advice. A recommendation can push the user toward one offer. Our tool kept the final choice with the person. It helped them see what the AI was likely to accept and how much room was left.


This distinction matters in systems where value can shift between sides. A helper can make one party better off by making the other worse off. We checked joint payoff, not only the tenant payoff, because a useful helper should not only move value from one side to the other.


Five actions for builders


  1. Add support once users must track more than about three things at once, and test your own limit.

    In our study, the sharp drop appeared after 3 issues. Your product may have a different limit.


  2. Show what is still possible and how close the user is, not what to do.

    The grid and progress bar helped people understand the deal space without taking over the offer.


  3. Keep the final choice with the user.

    The tool never told the tenant what to offer. It left the decision to the person.


  4. Check both sides’ payoffs when you test a helper.

    This helps you see whether the tool improves the deal or only shifts value to one side.


  5. Compare against a simpler display.

    This is our suggestion and was not tested in the study. A simpler display might help too, and it should be part of future comparisons.


The main takeaway is small but useful. When AI deal tools move beyond a few issues, people may need support that shows the possible deal space. The support should make the task clearer, while leaving the choice in human hands.


bottom of page