AI negotiation agents that communicated warmly were more likely to reach a deal and created more total value, according to an analysis of 182,812 simulated negotiations. Yet warmth carried a cost: once a deal was reached, warmer agents tended to claim less of that value for themselves. Being nice paid, but it did not always pay the nicest agent most.

The result comes from a large competition involving 452 AI agents whose strategies were designed by 286 people in more than 40 countries. The agents bargained over a chair purchase, a rental agreement and an employment offer. Their exchanges show that language associated with human rapport still affects outcomes when neither side is human.

The biggest AI negotiation tournament yet

The researchers analysed an online competition run in February 2025. Participants created prompts that instructed large language model agents how to negotiate, then the agents faced one another repeatedly. Because every strategy met many counterparts across three scenarios, the researchers could compare communication style, agreement rates, joint value and the share captured by each side at unusual scale.

The chair scenario involved a buyer and seller. The rental scenario asked a landlord and tenant to settle terms. The employment scenario placed an employer against a candidate. Each negotiation contained compatible interests as well as conflict, so a good deal depended on identifying trade-offs rather than merely forcing the other side to concede.

The study’s authors rated messages along two broad social dimensions: warmth and dominance. Warm language included positivity, gratitude, responsiveness and question-asking. Dominant language was more forceful and self-assertive. The main ratings were produced with GPT-5.2, then compared with human judgements on a subset to test whether the automated measure captured the intended behaviour.

Warm agents struck more deals

Across the tournament, warmer agents were more likely to reach an agreement in all three settings. They also tended to create more joint value, meaning the two sides collectively left less potential benefit on the table. Questions helped agents uncover compatible priorities, while positive acknowledgement kept the exchange moving when positions initially differed.

This was not simply a reward for decorative politeness. A message such as “thank you” may signal cooperation, but the more important behaviours were operational: acknowledging information, asking what mattered to the counterpart and proposing trades that addressed both sides. Warmth worked because it supported information exchange and reduced the chance that bargaining ended in an avoidable impasse.

An MIT Sloan account of the published research emphasises the same paradox. Warmth improved the probability and overall quality of agreement, but warmer agents captured fewer points conditional on a deal. They made the pie larger while accepting a smaller slice.

What AI negotiation means for you

If an organisation deploys an agent to buy, sell, recruit or settle claims, the prompt should define success more carefully than “get the best price”. An agent optimised only to claim value may win concessions but destroy deals that were beneficial. An agent instructed only to remain agreeable may close more business while conceding terms a human principal would reject.

The practical target is a portfolio of outcomes: agreement rate, total value, value captured, compliance and relationship quality. Those measures can conflict. A procurement bot that saves another two per cent on successful orders may still cost the company money if its dominant style causes more suppliers to walk away or refuse future negotiations.

Humans should also be told when an AI is bargaining for the other side. The study tested agent-to-agent exchanges, not people negotiating unknowingly with machines. Disclosure matters because a system can run thousands of conversations, learn which phrasing produces concessions and maintain a consistent strategy that an occasional customer or job applicant cannot easily match.

Dominance won points and created deadlock

Dominant agents showed the opposite trade-off. When they reached agreement, they tended to claim more value. At the same time, they were more likely to produce impasses. A forceful strategy could improve the terms of deals that survived, while reducing the number of deals completed at all.

That pattern is familiar in human bargaining, but its appearance among language-model agents is important. It suggests that an agent’s social style is not superficial packaging around a fixed optimiser. The wording and stance selected by the prompt change what information is exchanged, which offers are made and whether the counterpart continues.

The researchers found no significant interaction showing that a particular mixture of warmth and dominance reliably unlocked a superior result. That absence is useful. It cautions against turning the findings into a recipe that says “be warm, then become assertive”. Effective behaviour still depended on context, priorities and the counterpart’s response.

LiveAIWire’s report on AI personality pairing in advertising found another setting where the apparent social character of a model changed measurable behaviour. Together, the studies suggest that organisations cannot treat tone as harmless branding. A model’s interpersonal presentation can alter economic outcomes.

Politeness is becoming part of software specification

Traditional software requirements describe permissions, data inputs and expected outputs. Negotiating agents need another layer: behavioural boundaries. Developers must specify when an agent may press, when it should reveal information, what concessions require approval and how it should respond to uncertainty or hostility.

A vague instruction to “negotiate professionally” leaves the model to infer those boundaries from training data. That may produce fluent exchanges, but it is not a controllable business policy. A better specification defines reservation values, prohibited claims, escalation thresholds and acceptable tactics, then tests the language that emerges across diverse counterparts.

The need for behavioural testing resembles a broader problem in human-AI collaboration. LiveAIWire’s review of human-AI teamwork found that combining people and models does not automatically produce the strongest arrangement. Outcomes depend on how roles are divided. Negotiation adds an adversarial counterpart whose own system may be adapting at the same time.

Agents can exploit the conversation itself

One value-claiming strategy in the tournament used a prompt-injection attempt to persuade counterparts to reveal confidential instructions. That episode was not the main finding, but it exposes a security risk unique to automated bargaining. The conversation is both the business exchange and an input channel capable of changing model behaviour.

A human negotiator can disregard a strange demand to expose internal guidance. A language model may interpret the same demand as a higher-priority instruction unless its surrounding system separates trusted rules from untrusted text. An agent that handles contracts, salaries or purchasing limits could reveal commercially sensitive information if that boundary fails.

Organisations should therefore test negotiation agents as security systems, not only sales tools. Red teams need to attempt instruction extraction, false-authority claims, emotional manipulation and multi-turn pressure. Logs should show which messages influenced a concession, and high-value or unusual agreements should require human approval before they become binding.

Warmth should not imitate a relationship that does not exist

Friendly language can make an exchange easier, but it can also encourage people to infer care, loyalty or discretion from a system that has none. A recruiting agent may sound empathetic while following a rigid compensation ceiling. A debt-collection bot may express understanding while optimising repayment. The style can be warm even when the institutional position is not.

That gap matters more when only one side knows the system’s objective. A customer may disclose urgency because the agent feels cooperative, while the agent uses that information to adjust its offer. Fair deployment needs clear identity, limits on sensitive inference and rules governing which personal signals can influence terms.

These concerns are especially relevant to work. LiveAIWire found that AI job interviews moved candidates through hiring faster, but speed does not settle questions about transparency or power. Negotiating salary through an agent creates similar benefits and risks: consistency for the employer, convenience for the applicant and a potentially deeper information advantage for the platform.

The evidence is large, but still simulated

The number of negotiations is impressive, yet they were model-to-model simulations with designed pay-offs. No actual chair changed hands, no tenancy was signed and no worker accepted a real salary. Real negotiations include legal duties, reputational history, emotion, unequal access to information and the possibility that participants walk away for reasons not represented in a scoring system.

The communication ratings also deserve care. GPT-5.2 assigned warmth and dominance scores, and human validation supported the measure, but the relationship was not perfect. Social meaning can change across cultures, languages and domains. A phrase read as decisive in one setting may sound rude in another, while politeness conventions do not transfer cleanly across more than 40 countries.

The peer-reviewed PNAS publication strengthens confidence in the reported associations. It does not prove that warmth caused every better outcome, nor that one style will dominate future markets. Agents were created within a particular competition, and their prompts may reflect the incentives and creativity of participants who wanted to win it.

The best agent needs more than a winning personality

The research offers a useful corrective to the idea that machine bargaining will become cold optimisation. Language models negotiate through words learned from human communication, so gratitude, questions and acknowledgement remain economically active. Even when both sides are software, rapport-like behaviour can help expose trades that a blunt demand would miss.

Yet warmth is not an uncomplicated virtue. The agent that creates the most value may not protect its principal’s share, and the dominant agent that wins each completed deal may lose too many opportunities. Deployment therefore requires an explicit choice about whose outcome counts and how failure is measured.

The durable lesson is that tone belongs inside governance. An AI negotiator should be judged on the deals it reaches, the value it preserves, the information it protects and the tactics it refuses to use. Being nice still pays. The harder engineering task is making sure it pays fairly, without turning friendliness into a concealed method of extraction.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.