AI agents can negotiate, and the published research shows they reach deals, but how well they negotiate depends on the role they play, the structure around them and the counterparty across the table. Across studies from Facebook AI Research's 2017 bargaining models to 2025 marketplace simulations, the consistent finding is wide variation. Some agents close sound deals, while others overspend, give way under pressure or break deals that should have closed.

TL;DR: Language models can reach deals in simple price haggling but struggle in multi-issue, cooperative and adversarial settings. The buyer seat is measurably harder than the seller seat, and in one ACL 2024 study a deterministic offer engine helped buyers where larger models did not. Emotional appeals, response speed and prompt injection all moved outcomes in published experiments, so media teams should judge the terms an agent signs, not just whether it finished.

Can AI agents negotiate well?

Whether AI agents negotiate well depends on the setting. Davidson and colleagues at EPFL, in a paper accepted to ICLR 2024, found that "even the most powerful models sometimes 'lose' to weaker opponents" (Davidson et al., 2024). Zhu et al. described AI-mediated deal-making as "an inherently imbalanced game" in which different agents secure significantly different outcomes for their users (Zhu et al., 2025).

How has AI negotiation been tested in advertising?

Most academic studies use generic games: splitting items, haggling over a product price or trading resources. Media deals add terms beyond price, such as make-goods, viewability floors and delivery guarantees, which is where a domain benchmark adds something.

Concourse Bench v1, built by Alkimi, is one worked example. It gave eight model-driven buyers a fictional media brief and one budget, and had each negotiate with three sellers controlled by fixed software policies, so it tested the buyer side only. Across 96 registered attempts, 55 ended in a completed buy, and 95 of the 110 signed contracts promised a viewability floor below 70%.

The agents closed deals, and many of the terms left the buyer weakly protected. Completion and deal quality need to be measured separately.

Which studies have tested how well language models negotiate?

Ten studies, oldest first, each with its method in one line and its headline finding.

1. Deal or No Deal? (Lewis et al., Facebook AI Research, 2017)

Method: models trained end to end on a large dataset of human-to-human negotiations over a multi-issue bargaining task.

Finding: the authors showed for the first time that models could learn to negotiate without annotated dialogue states. A technique called dialogue rollouts, in which the model simulates complete continuations of the conversation before replying, "dramatically improves performance" (arXiv).

2. Self-play and AI feedback (Fu, Peng, Khot and Lapata, 2023)

Method: two language models played buyer and seller over repeated rounds, with a third model acting as a critic that suggested better strategies.

Finding: only some models improved their deal price from AI feedback, while weaker ones misread the rules or could not use it. Stronger agents improved over multiple rounds "yet have a higher risk of breaking the deal" (arXiv).

3. LLM stakeholder negotiation (Abdelnabi et al., 2023)

Method: multi-agent, multi-issue negotiation games, including games with greedy and adversarial players.

Finding: GPT-3.5 and small models "mostly fail", and GPT-4 and large models such as Llama-3 70b "still underperform" on the benchmark (arXiv)

4. Evaluating language model agency (Davidson et al., ICLR 2024)

Method: six widely used models played negotiation games against copies of themselves and against each other.

Finding: only the closed-source models tested could complete the tasks, and cooperative bargaining games were the hardest. Even the most capable models sometimes lost to weaker opponents (arXiv).

5. NegotiationArena (Bianchi et al., Stanford University, 2024)

Method: ultimatum games, trading games and buy and sell price negotiations between language model agents over multiple turns.

Finding: behavioural tactics moved results, and "by pretending to be desolate and desperate", agents improved their payoffs by 20% when negotiating against standard GPT-4. The authors also catalogued irrational negotiation behaviours, many of which also appear in humans (arXiv).

6. Measuring bargaining abilities (Xia et al., ACL 2024 Findings)

Method: bargaining modelled as an asymmetric game with incomplete information, using real product prices from a dataset called AmazonHistoryPrice.

Finding: playing the buyer is "much harder" than playing the seller, and increasing model size did not effectively improve buyer performance. Pairing a deterministic offer generator with a language model that only phrases the offers raised buyer deal rates from 26.67% to 88.88% (arXiv).

7. International AI Negotiation Competition (Vaccaro et al., 2025)

Method: competition entrants designed prompts for negotiation agents, which then ran more than 180,000 negotiations.

Finding: warmth was "consistently associated with superior outcomes across all key performance metrics", while dominant agents were especially effective at claiming value. Longer conversations were strongly associated with impasses, and the authors observed AI-specific tactics including prompt injection (arXiv).

8. The Automated but Risky Game (Zhu et al., 2025)

Method: researchers from the University of Toronto, MIT, Stanford University and other institutions had consumer and merchant agents fully automate negotiations and transactions.

Finding: different agents achieved significantly different outcomes, and behavioural anomalies caused losses on both sides, "such as overspending or accepting unreasonable deals". The authors concluded: "Users should exercise caution when delegating business decisions to AI agents" (arXiv).

9. Project Vend (Anthropic and Andon Labs, 2025)

Method: Claude Sonnet 3.7 ran a small automated shop in Anthropic's San Francisco office for about a month, setting prices and dealing with customers over Slack.

Finding: the agent found suppliers for specialty items and refused sensitive requests, but it was "cajoled via Slack messages into providing numerous discount codes" and gave some items away. Anthropic concluded it made too many mistakes to run the shop successfully, while seeing clear paths to improvement (Anthropic).

10. Magentic Marketplace (Bansal et al., Microsoft, 2025)

Method: an open-source simulated two-sided market in which assistant agents represent consumers and service agents represent competing businesses.

Finding: frontier models "can approach optimal welfare", but only under ideal search conditions, and performance degraded sharply with scale. All models showed "severe first-proposal bias", giving response speed a 10 to 30 times advantage over quality (arXiv).

What patterns run across the research?

Five patterns recur across the ten studies.

The buyer seat is harder. Xia et al. measured the gap directly, and Fu et al. found one model improved more easily as a seller than as a buyer.
Structure matters as much as model size. Dialogue rollouts in 2017 and the offer generator in 2024 both added a planning or pricing component the language model lacked.
The counterparty shapes the result. Strong models lost to weaker ones in Davidson et al., and in Magentic Marketplace the first offer won out of proportion to its quality.
Persuasion works on machines. Desperation, warmth and persistent requests all moved outcomes, and prompt injection appeared as a competition tactic.
Finishing is not the same as winning. Stronger agents in Fu et al. ran a higher risk of breaking deals, and Zhu et al. recorded overspending and acceptance of unreasonable terms.

What does the research mean for media deals?

Media deals combine the hardest conditions in these studies: several issues at once, a buyer seat, incomplete information and terms that matter long after the conversation ends. An agent that pushes a price down while conceding the viewability floor has completed a negotiation and still lost the deal.

First, test a buying agent against firm and adversarial counterparties before any money moves. Results against one opponent say little about another, as explored in what a multi-round AI negotiation looks like and in what happens when two AI agents cannot agree. Teams ready to see how their own buying agent handles these conditions can scope a Concourse benchmark run.

Second, score the contract rather than the conversation. Price, protections and budget adherence all belong in the evaluation, and the software around the model shapes much of that, as set out in why the buyer harness matters as much as the model.

Third, agree on structured terms rather than conversational assent. IAB Tech Lab describes AAMP as an initiative for bringing agentic AI into advertising "in a responsible and scalable fashion", built on "the standards and taxonomies that power that industry", and AgenticAdvertising.org describes AdCP as an open protocol that lets agents discover, plan, buy, sell and measure media. Structured terms narrow the space in which a warm tone or a desperate plea can shift the outcome.

The research also supports granting authority in stages. Earned autonomy describes five steps: observe, recommend, draft, human-approved action and bounded automatic action, with each step granted on evidence from the one before.

Where does this leave buyers and sellers?

The research does not say AI agents cannot negotiate. It says their ability is uneven, sensitive to tactics and easy to overstate when completion is the only measure.

At Alkimi, that is the reasoning behind testing before trading. Concourse benchmarks AI buyers on whether they complete a brief and on the protections in the contracts they sign, with a second version now running, and the Gym lets buyers and sellers build their own agents, set the rules they work to and spar against counterparty agents before any real money moves. To test a specific agent, use the Concourse scoping form.

Frequently asked questions

Can AI agents negotiate better than humans?

The studies in this roundup mostly compare models with each other rather than with human negotiators, so they do not settle that question. Bianchi et al. did find that many irrational behaviours seen in language model negotiators also appear in humans.

Why is buying harder than selling for an AI agent?

Xia et al. measured the gap on real product prices and closed much of it by letting a deterministic generator set the buyer's offers. That suggests buyer agents struggle most with deciding what to offer, not with how to phrase it.

Can an AI negotiator be manipulated?

Published experiments show it can, at least in test settings. Emotional appeals, persistent requests and prompt injection all changed outcomes in the studies above.

How should an advertiser test an AI negotiating agent?

Run it against several counterparties, including firm and adversarial ones, and record every term it agrees. Judge the contract it signs as well as whether it finished.

Do any of these studies cover media buying?

The academic studies use general bargaining tasks, consumer goods and simulated marketplaces. Concourse Bench v1 is a media-specific benchmark that tested AI buyers against fixed-policy sellers on a fictional brief.

Can AI Negotiate? What the Research Says About Language Models at the Bargaining Table