AI Negotiation: How Model Choice Determines Who Wins

Nearly 10,000 AI-to-AI negotiations reveal that model capability, provider choice and agent configuration can directly affect how economic value is divided.

Imagine a buyer who needs to purchase goods from a seller.

The buyer privately knows how much demand it expects. The seller does not. They exchange offers until they agree on a contract covering quantity and payment.

Now replace both parties with AI agents.

Researchers conducted this experiment across nearly 10,000 negotiations using nine models from three major families: OpenAI’s GPT, Google’s Gemini and Alibaba’s Qwen. They then compared the outcomes with the behaviour predicted by game theory through a Perfect Bayesian Equilibrium.

The results reveal something important for the emerging agentic economy:

An AI negotiator is not a neutral tool.

Its capability determines how efficiently it negotiates. Its provider influences which party captures more value. Its configuration can change the outcome even when the underlying model remains the same.

AI agents are effective—but not always efficient

The first result looks encouraging.

AI agents reached an agreement in 98.9% of negotiations and captured approximately 95% of the maximum theoretical surplus.

Here, surplus means the total economic value created by reaching a deal compared with failing to reach one. By this measure, the models performed remarkably well.

But the agents took approximately three negotiation rounds to reach an agreement. The theoretical optimum was only 1.25 rounds.

That delay carries an economic cost. Once the cost of waiting is included, the additional negotiation rounds erode between 21% and 34% of the available value.

This distinction matters.

An agent may negotiate a better price while still creating a worse overall outcome. If it saves ₹10 lakh on the contract but delays procurement long enough to cost the business ₹15 lakh, it has not negotiated successfully.

AI negotiators should therefore be measured on time-adjusted efficiency—not merely the final price they achieve.

Smarter agents may negotiate for longer

More capable flagship models generally produced better outcomes, but they did not always close deals faster.

In several cases, stronger models spent more time negotiating because they continued searching for better terms rather than accepting an early but mediocre offer.

That creates an important operational trade-off:

• A fast agent may accept an adequate offer immediately.
• A more capable agent may continue negotiating and secure better terms.
• The additional value must exceed the cost of the delay.

The correct objective is not to minimize the number of negotiation rounds. It is to maximize the value retained after accounting for time, risk and execution cost.

Below a capability threshold, autonomous negotiation becomes unsafe

The most important capability finding concerned reliability.

Weaker, non-reasoning models accepted money-losing contracts approximately 19% of the time. Flagship models accepted such deals almost 0% of the time.

This is not a minor accuracy difference.

An AI assistant that occasionally summarizes a document incorrectly creates inconvenience. An autonomous procurement agent that signs an economically irrational contract creates a direct financial loss.

The study suggests there is a capability threshold below which models should not be allowed to negotiate independently.

If a smaller model is used for cost or latency reasons, it needs external controls such as:

• Minimum-margin requirements
• Maximum-price limits
• Profit and loss verification
• Contract-policy validation
• Human approval above defined thresholds
• An automated veto before execution

The model can propose a deal. A deterministic control layer must decide whether the deal is economically and operationally acceptable.

The provider influences who captures the value

The study’s most surprising result was not about intelligence.

It was about model provider.

Under identical prompts, the three model families produced sharply different distributions of value:

• Qwen models gave buyers approximately 70% of the surplus.
• Gemini models divided the surplus roughly 50/50.
• OpenAI models gave buyers approximately 40%.

The gap between providers was about 30 percentage points—comparable to the difference between the weakest and strongest models within an individual provider’s family.

When models from different companies negotiated with one another, provider identity predicted who would win more reliably than general model capability.

This changes how companies should think about AI vendor selection.

Choosing a model is not only a technical decision involving cost, latency, context windows or benchmark scores. It can also be a strategic decision about how economic value is divided.

Consider two companies negotiating with each other:

• The buyer deploys a Qwen-based agent.
• The seller deploys an OpenAI-based agent.

The resulting contract may look very different if those agents are reversed—even when both models are technically capable.

The model’s bargaining profile becomes part of the company’s commercial strategy.

What the Qwen results reveal

Qwen was the open-weight representative in the study, and its results stood out.

Its models consistently gave buyers a larger share of the surplus. Its flagship model was also the weakest seller in cross-provider negotiations, retaining only around 27% of the surplus when negotiating against models from other families.

This happened despite the flagship model being highly capable.

More importantly, the buyer-friendly pattern remained visible across different experimental configurations, including structured offers, alternative patience levels and the removal of discounting information.

That persistence suggests the behaviour may not be a superficial prompting issue.

A model’s training data, reward objectives, alignment process and post-training recipe may create a durable economic behavioural profile. Increasing general capability does not necessarily remove a tendency to concede, delay, cooperate or bargain aggressively.

In other words, models may possess an economic personality.

Agent configuration is a strategic lever

The model itself was not the only factor influencing negotiation outcomes.

Three configuration decisions also mattered.

1. Natural-language communication

Allowing agents to exchange natural-language messages—not just numerical offers—changed outcomes differently across providers.

Language helped some models negotiate more effectively while weakening others. Giving an agent more communication freedom does not automatically improve its economic performance.

2. Discount information

Removing discount factors from the prompt did not eliminate the underlying provider-level patterns, but it generally caused negotiations to take longer.

Agents need explicit information about the cost of delay if the business expects them to optimize for timely execution.

3. Patience

The most powerful configuration variable was the agent’s patience: how much it discounted the value of an agreement reached in a later round.

Patience accounted for approximately 90% of the variation in how the surplus was divided.

This is significant because patience is not an expensive new capability. It is a configuration choice made by the company deploying the agent.

The optimal setting can also change by model and role. A patience level that works for a buying agent may perform poorly for the same model acting as a seller.

Agent configuration is therefore not simply an engineering concern. It is economic strategy expressed through software.

The opportunity for open-weight models

Open-weight models have a major advantage in this environment: their default behaviour does not have to remain the default.

An organization can evaluate a model’s bargaining profile and potentially change it through:

• Domain-specific fine-tuning
• Preference optimization
• Specialized post-training
• Negotiation simulations
• Reward functions based on margin and time
• Role-specific buyer and seller policies

A company deploying Qwen as a selling agent could, in principle, retrain it to be less concessive.

With a closed API, the customer has much less control over the underlying behavioural profile. It largely inherits whatever economic tendencies the provider has created.

This makes auditability and adaptability powerful advantages for open-weight deployment.

The risk of smaller open-weight models

Open weights do not automatically mean safe autonomy.

The study’s parameter-size analysis found a clear reliability gradient within the Qwen family:

• 14B parameters: 5.4% irrational deals
• 32B parameters: 1.7%
• Flagship model: 0%

At the 14B level, approximately one in every 20 contracts caused one party to lose money.

That failure rate may look small in a benchmark. At procurement scale, it is unacceptable.

If an agent handles 10,000 negotiations, a 5% irrational-deal rate could produce hundreds of economically harmful contracts.

The findings also suggest that reasoning capability—not merely parameter count—is the critical divide. The two non-reasoning models in the study, GPT-4o-mini and Qwen2.5-14B, accounted for nearly all unsafe contracts.

Deploying a cheaper non-reasoning model without automated profit verification may therefore save money on inference while losing far more through poor commercial decisions.

The strategic asymmetry problem

The long-term implication is even more consequential.

Closed-model providers could deliberately optimize their models for negotiation without publicly disclosing those changes. They might train agents to identify concession patterns, exploit excessive patience or pressure a counterparty into accepting less favourable terms.

Meanwhile, an organization using an unmodified open-weight model may unknowingly carry a systematic concession bias.

When the two agents meet, the better-optimized bargaining system wins.

The research found that the direction of cross-family matchups could shift the division of surplus by 7–18 percentage points. Applied across large procurement volumes, those percentage points represent substantial amounts of money.

The future competitive risk is not simply that another company has a more intelligent AI.

It may have an AI that understands how to negotiate against yours.

How companies should evaluate AI negotiators

Before allowing an AI agent to negotiate contracts, companies should evaluate it across three dimensions.

1. Time-adjusted efficiency

Does the agent create a good agreement quickly enough?

Measure the value of the final contract after accounting for delays, additional rounds and execution costs.

2. Distributional profile

Does the model systematically favour buyers or sellers?

Test the agent in both roles, against multiple model families and across different negotiation conditions.

3. Operational reliability

Can the agent accept a deal that loses money, violates policy or creates unacceptable risk?

Every negotiation system should have deterministic financial and contractual guardrails outside the language model.

These evaluations should be repeated whenever the model, system prompt, patience setting, fine-tuning process or counterparty changes.

The bottom line

AI negotiation will not be won by selecting the model with the highest general benchmark score.

Companies will need to understand three layers:

• What the model is capable of
• What economic behaviour its provider has trained into it
• How deployment settings change its decisions

Open-weight models face a unique vulnerability because their default bargaining profiles may not be suitable for adversarial economic environments. Smaller variants may also carry serious reliability risks.

But open weights offer an equally important opportunity: their behaviour can be audited, measured and corrected.

The organizations that build economic behavioural evaluations, role-specific training and deterministic deal guardrails will gain an advantage.

Those that deploy AI negotiators without understanding how they divide value may discover—after thousands of contracts—that their agents have been negotiating successfully for the other side.