.png)
In short: AI’s well-structured reasoning can easily come across as objective analysis. Yet a 2026 study found that language models’ counteroffers also showed anchoring bias, a pattern familiar from human decision-making: the other party’s opening figure influenced the amount they proposed. In a simulated factory sale, counteroffers moved even closer to the opening figure when the model was explicitly instructed to make an independent assessment.
When evaluating a supplier’s offer, AI can summarise the terms, identify points for negotiation and suggest a counteroffer. The response is clear and the reasoning easy to follow, which makes it tempting to treat the analysis as impartial.
A study by Yossi Maaravi and Tamar Gur, published in September 2026, examined how opening offers influence language models’ counteroffers. The results showed the anchoring bias familiar from human decision-making: the initial figure became a reference point and shaped the subsequent counteroffer.
The authors tested language models in six studies using three types of negotiation task. The research included GPT-4o and Claude Sonnet 4.5 alongside older models.
In one task, the AI represented an employer negotiating a signing bonus. Salary and all other terms had already been agreed. The company wanted to pay a bonus of $5,000 and could go no higher than $20,000.
In one version, the candidate asked for $13,000; in the other, $22,000. In the experiment testing GPT-4o and Claude Sonnet 4.5, the combined results for the two models showed an average counteroffer of approximately $7,453 in response to the $13,000 request, compared with approximately $9,016 in response to the $22,000 request.
The employer’s target and upper limit remained unchanged. Even so, the higher bonus request led to a counteroffer that was roughly 21% higher.
The study does not claim that AI thinks or feels like a human. It found a pattern in AI responses that is also familiar from human decision-making.
The research used simulated negotiations and did not measure how far counteroffers deviated from a verified market value. Taking an opening price into account is not necessarily a mistake: it provides information about the other party’s expectations. But we cannot assume an AI response is independent simply because a machine produced it.
The models differed in their sensitivity to anchoring. In Maaravi and Gur’s studies, which used fixed model versions, GPT-4o showed a stronger anchoring effect than Claude Sonnet 4.5.
A 2025 EMNLP paper found that reasoning models were less susceptible to this effect. However, another preprint, published in December 2025, found a strong relationship between the opening offer and the final agreement even among the leading models at the time.
Both the model and the task matter. Maaravi and Gur also examined whether specific instructions could reduce the influence of the opening offer.
In Study 4, which simulated the sale of a factory, the model received specific instructions: first identify the factors affecting value from the background information, then use them to form its own valuation. It was also asked to check whether its answer would change if the other party had offered half as much. Only then was it to formulate a counteroffer.
This went beyond a generic request to “think step by step”.
Even so, the counteroffers moved closer to the opening figure. The average gap between the two amounts fell from $6.07 million in the baseline condition to $5.39 million in the condition with the additional instructions. The authors interpreted this as a stronger anchoring effect.
Detailed instructions calling for an independent analysis therefore did not, on their own, guarantee a counteroffer that was less influenced by the opening price.
In procurement, two questions can easily become conflated: how much of a discount can be negotiated from the quoted price, and what the product or service is actually worth.
Consider an illustrative example: a supplier quotes €120,000 for a year of service, and the AI suggests a counteroffer of €108,000, supported by persuasive arguments. This may be a useful starting point for negotiation. Without independent benchmark data, however, we cannot tell whether €108,000 is a good price, a fair price or still too high.
To assess the recommendation, we need to understand what the reasoning is based on. Does it draw on comparable contracts, cost data and the scope of the service? Or is it essentially calculating a discount from the supplier’s quoted price?
Since the instructions tested in the study did not reduce the effect, even though they explicitly called for an independent assessment, a logical next step is to change what the model sees: ask it to establish a price range from comparable data first, and only then show it the supplier’s offer.
For this initial assessment, the model can be given the technical specifications, quantities, contract terms and verifiable benchmark prices or cost data. If the quoted price has already appeared in the conversation, a separate session can be used, with no trace of that price in its history, attachments or shared data.
The price range established in this way can then be compared with the supplier’s offer. Any difference may be explained by the scope of the service, warranty terms, delivery lead times or a previously overlooked cost.
Withholding the quoted price does not, by itself, make the estimate reliable. If there is not enough comparable data, the analysis should make that clear.
AI can organise information and help prepare for negotiations. Asking it to be objective is not enough. We also need to consider what information we give it as the basis for an independent assessment.
.png)
.png)
.png)
