An award-winning AIT study reveals when negotiations with AI become too complex for people and introduces a visual tool that helps them negotiate more effectively while remaining in control.
Imagine negotiating your next apartment lease. Rent, deposit, lease length, and whether you can keep a pet. The party across the table, however, is not a person. It is an artificial intelligence system capable of considering every issue simultaneously, evaluating hundreds of counter-offers per second, and quietly inferring what you care about from the pattern of your bids.
This is no longer hypothetical. Walmart already uses an AI agent to negotiate supplier contracts, Maersk settles shipping rates the same way, and legal-aid systems draft settlements overnight. What nobody had measured is how well humans perform in these negotiations.
Mehul Parmar, a PhD candidate in Information Management at the Faculty of Advanced Science and Technology, AIT, set out to answer this question with his advisor Dr Chaklam Silpasuwanchai, Assistant Professor, Computer Science and Communication Technology (CSCT).
Their paper From Overload to Convergence: Supporting Multi-Issue Human-AI Negotiation with Bayesian Visualization, was presented at CHI 2026, a leading international conference in human-computer interaction with over 6000 submissions globally, where it received a Best Paper Award (5% percentile).
Three issues, and then a cliff
Negotiation researchers have long argued about the effects of complexity. One camp argues that having more issues on the table creates more room for creative trade-offs and better deals for everyone. Another argues that additional issues simply overwhelm the human mind. Neither side, however, had tested the question against an AI opponent, which does not tire or forget.
The AIT researchers developed a property rental negotiation scenario and asked 32 participants to complete it four times, with one, three, five and seven issues on the table. Every participant played the tenant, while an AI agent played the landlord. The AI worked from its own hidden scoring system and was instructed to maximise its own outcome. Because neither side could see the other’s scoring, a good deal had to be discovered through conversation rather than calculated in advance.
The results were clear enough to deserve a name. With up to three issues, participants were able to hold their own. At five and seven issues, however, their performance fell sharply. They reached worse deals, paused longer before responding, and repeatedly circled back to offers that had already been rejected. The authors call this the plateau-cliff effect.
“As AI systems increasingly mediate our negotiations, we must ensure technology empowers rather than overwhelms. Our latest research at AIT reveals a clear cognitive ceiling: without support, human performance drops sharply after juggling three issues. By introducing a novel Bayesian visualization to map the narrowing space of agreement, we can preserve human control and secure better outcomes in complex human-AI interactions,” said Dr Chaklam.
Tellingly, participants’ frustration increased gradually, while their performance declined suddenly. They were not giving up. They were running into a genuine ceiling on how many interlocking trade-offs a person can hold in mind at once.
A pair of glasses, not an autopilot
Having located the ceiling, the authors built something to sit beneath it.
The tool has two main components. First, a live grid shows where a deal is still possible, highlighting the options the AI appears willing to accept that are genuinely good for the user. Second, a single horizontal bar summarises the negotiation at a glance. It shows how much uncertainty remains, and whether the deal is leaning toward or away from the users interests. Behind both, the system revises its estimate of the AI’s preferences after each offer, and shows how confident it is in that estimate.
The design approach is as important as mathematics. The tool demonstrates rather than instructs. It does not tell the user to “offer $1,900”; but converts the act of remembering into one of observation.
With the tool, the cliff disappeared. Participants’ payoffs held steady across all four levels of complexity. Negotiations took fewer turns, participants replied faster, and they reported lower mental effort and greater confidence. They also rated the interface 81 on the standard usability scale, placing it comfortably in the “excellent” band.
The most important result, however, was what did not happen. The tool did not take value away from the AI and hand it to the human. In negotiations involving seven issues, agreements moved measurably closer to the theoretically best outcome available, recovering value that both parties had been leaving on the table.
Beyond the lease
The findings extend far beyond rental agreements. Supplier and procurement negotiations, insurance settlements, and online dispute resolution all involve multiple interlocking terms and, increasingly, an automated counterpart. The underlying design principles reach even further: from product recommenders that bury shoppers in criteria to medical decision aids where patients weigh treatment options to financial planning tools.
Four principles carry across. Switch support on only once complexity begins to bite. Separate “which options are still acceptable” from “how close are we to agreement.” Assist without prescribing. Remove waste rather than redistribute gains, because people accept help that clears away inefficiency but not help that quietly moves value from one side to the other.
The authors gather these under a principle they call Cognitive Harmony: extend human capacity without eroding human agency. As AI moves into high-stakes, many-sided decisions, the prescription is compact. Find the cliff, support people at it, and leave the judgement where it belongs.
Read the full paper: From Overload to Convergence: Supporting Multi-Issue Human-AI Negotiation with Bayesian Visualization
Edited by: Office of Communications and Public Affairs
