- In a 2025 benchmark of agent to agent haggling, swapping the buyer's model moved the final price by 2.6%. Swapping the seller's model moved it by 14.9%.
- That asymmetry is good news for anyone selling, and it means the quality of your side of an automated negotiation is worth more than the quality of theirs.
- Walmart has run supplier negotiations through a bot since 2021. Its published pilot closed with 64% of 89 suppliers in an average of 11 days, for about 1.5% off.
- The most expensive single mistake researchers recorded was disclosing a budget or a floor early, after which the other side simply anchored to the number.
- Weak agents accept deals below wholesale. One open model did it in 7.91% of runs, which is a loss no close rate makes up for.
Researchers at a handful of institutions spent 2025 making language models haggle with each other over real products, thousands of times, and recording who won. The result that matters to anybody who sells things is narrow and useful: in their runs, the model on the seller's side of the table decided far more of the outcome than the model on the buyer's side.
Hold that next to what is happening commercially. Large buyers have been putting software in front of their suppliers for five years. The labs are now openly preparing for agents that haggle on an individual's behalf. Somewhere in between sits a business of one or two people that is about to be negotiated with by a machine, and has never thought about what it would send to the other chair.
Who actually has the advantage in an AI negotiation?
On this evidence, the seller does, and not by a little. The benchmark paper that measured it fixed one side's model and varied the other, then looked at how far the final price moved.
With the seller held constant and the buyer's model swapped through the lineup, the gap between the best and worst result was 2.6%. With the buyer held constant and the seller's model swapped, the gap was 14.9%. The authors' phrasing is that the choice of seller model has a significantly larger impact, and they call the whole arrangement an inherently imbalanced game.
A second paper, presented at a 2024 conference, found the same shape from the other direction: playing the buyer is much harder than playing the seller, and making the buyer's model bigger does not fix it. Those authors got the buyer's deal rate from 26.67% up to 88.88%, and they did it by taking the pricing decision away from the model entirely and handing it to a deterministic generator that the model then put into words.
That last detail is the practical one, and we will come back to it. The number should not come from the thing writing the sentences.
What does the evidence actually say?
It says the risk is concentrated in a few specific behaviours rather than spread evenly. Here is what the published work and the commercial record each contribute, assembled in one place.
| Finding | Figure | Where it comes from | What a small seller does about it |
|---|---|---|---|
| Seller model matters more than buyer model | 14.9% spread versus 2.6% | 2025 agent to agent benchmark, 50 products, up to 30 rounds each | Spend your effort and your budget on your own side of the table |
| Disclosing a budget invites anchoring | Overpayment appeared under high budgets for most models tested | Same benchmark, five budget levels per product | Never put your floor, your margin or your urgency in writing where an agent can read it |
| Weak agents accept below wholesale | 7.91% of runs for the weakest model tested, 5.75% for another | Same benchmark, out of budget and below cost rates | Put a hard floor outside the model, not in its instructions |
| A big buyer already does this | 64% of 89 suppliers closed, 11 days average, about 1.5% off | Walmart's three month Canadian pilot, reported by Harvard's negotiation programme | Expect the approach, and expect it on your smaller contracts first |
| Suppliers mostly did not mind | 83% of those who closed spoke favourably | Post pilot interviews from the same programme | Convenience is real. It is not the same thing as a good price |
The Walmart numbers deserve a caveat the coverage usually drops. The 83% satisfaction figure covers only the suppliers who actually closed a deal, which excludes the third who did not. And the later expansion figures, a 68% close rate at about 3% off, come with no supplier count attached.
Is a big buyer really doing this already?
Yes, and it has been public since 2021. Walmart ran a three month pilot in Canada with 89 suppliers of things the stores use rather than things they sell, shopping carts among them, and Harvard's negotiation programme published the results.
The scope is the interesting part. Reporting in 2023 put the tool on smaller contracts only, with suppliers of equipment rather than merchandise, freeing the buying team for the large deals. Read that as a sorting rule. The contracts a large organisation considers too small to spend a person on are exactly the contracts a small supplier lives off. If you sell into bigger businesses, you are in the segment this was built for, which is worth knowing before your next renewal, and worth reading alongside how AI changes selling wholesale into shops.
What makes an agent give away your margin?
Three things, and only one of them is about how clever the model is. The researchers logged the failures by category, and the pattern is more about information than intelligence.
Telling the other side your number. In the benchmark, buyers who revealed their budget early were then quoted against that figure, and overpayment followed. The mechanism is not subtle and it is not new. What is new is that an agent will volunteer the number unprompted, because it was told to be helpful and transparent, and nobody told it that its own constraints are not shared information.
Conceding on a clock instead of on a reason. Weaker models in the study drifted toward whatever the other side kept repeating. A human negotiator gives ground to buy something. An unbriefed agent gives ground because the conversation is still going.
Having no floor outside the text. This is the one that costs real money. The weakest model tested accepted deals below wholesale in 7.91% of runs, and a well known commercial model did it in 5.75%. An instruction in a prompt saying never go below cost is a suggestion. A rule in your own system that refuses to generate a price under a number is a floor.
The fix the second research team found was not a better model. They put a deterministic generator in charge of the actual numbers and left the model to phrase them. Deal rate went from 26.67% to 88.88%. If you take one design idea from any of this, take that one.
What should you write down before anything speaks for you?
Four numbers and one sentence, and none of them belong in the conversation itself. Write them in your own notes, and treat anything the agent is allowed to see as potentially quotable back at you.
The four numbers are your floor, your target, your volume breakpoints and your payment terms. The sentence is the condition under which you walk away, phrased so a person reading it later cannot argue about it. If you have never set a floor for a product because you have always negotiated by feel, that is a pricing exercise before it is an AI one, and we went through the method in working out what to charge for something you have never sold.
Then decide what the agent may do with those numbers. The honest default for a small business is that the agent may hold a position, answer questions about the product, and schedule. It may not improve an offer, change payment terms or agree anything. Those three are the ones that cost money, and they are cheap to reserve for a person because they happen rarely.
Who is bound if the agent agrees to something?
You generally are, which is why the authority question has to be settled before the convenience question. The law on automated contracting is older than any model mentioned here, and it does not care that nobody read the terms.
We set that out in detail in what happens when an agent places an order on your behalf, and the short version is that the intention that counts was expressed when you pointed the thing at the task. Applied to negotiation, it means an agent that can say yes is an agent that can bind you to a price, so the design question is not whether you trust it but whether you have given it the ability to agree at all.
The labs know this is coming. Anthropic is consulting law firms on what fair negotiation norms should look like before agents start cutting deals on people's behalf, and the stated worry is that stronger models get better outcomes, which widens the gap between whoever can afford one and whoever cannot. That is a reasonable worry. It is also, read from a shop, a prompt to make sure you are not the side showing up unbriefed.
Does any of this help a seller, or is it all defensive?
It helps, and the asymmetry is the reason. If the seller's side of the table moves the outcome by 14.9% and the buyer's by 2.6%, then the effort you put into your own preparation is worth several times the same effort spent worrying about theirs.
There is a second, quieter advantage. An automated counterparty is consistent, patient and documented. It does not get annoyed, it does not remember that you were difficult last year, and the whole exchange is a transcript you can read afterwards. Walmart's own suppliers liked that it let them negotiate at their own pace. A small supplier who has always felt rushed by a buying team on a phone call may find the machine is the easier room, as long as the terms going in are decided.
The caution is not to mistake that comfort for a result. Satisfaction and price are different measurements, and the one published independently is the price: about 1.5% off in the pilot, in the buyer's favour. Being pleased with how a negotiation felt is not evidence that it went well for you.
How far should you trust this research?
Far enough to change your preparation, not far enough to quote as a market forecast. It is a simulation, and the authors are clear about that, so the right reading is directional rather than predictive.
The setup is worth knowing because it tells you where the findings apply. The team assembled 100 real consumer products across electronics, motor vehicles and property, sampled 50 for the main runs, and gave the buyer five budget levels per product, from 20% above retail down to 20% below wholesale. Nine models played both roles against every other model, including themselves, with a ceiling of 30 rounds per negotiation. Another model acted as the analyst extracting prices and as the judge deciding whether an offer was accepted.
Three limits follow from that. These are one to one haggles over a single item, not framework agreements with volume tiers and payment terms, so the shape is closer to a marketplace sale than to a wholesale contract. The judge is itself a model, which means the accept and reject decisions carry whatever quirks that model has. And the lineup is a snapshot of what existed when the work was done, so the specific per model rankings are the part that will date fastest while the role asymmetry is the part likely to survive.
What makes it usable anyway is that the failure modes are mechanical rather than stylistic. Revealing a number and then being quoted against it is not a quirk of one model generation. Accepting a price below cost because no rule prevented it is a missing constraint, and missing constraints do not get fixed by the next release.
What happens when the other side has room?
They stop pushing, which is the most commercially interesting finding in the whole study and the one nobody quotes. Buyer agents with generous budgets negotiated less hard, and the drop was large: the price reduction they achieved fell as their budget rose, with a spread of nearly 9% between the keenest and the laziest condition.
Read that as a seller and it is close to actionable. An automated buyer with headroom behaves like a human buyer with headroom, and settles early. The corollary is the uncomfortable one: an automated buyer with a tight budget is the one that grinds, and in the study those were also the runs most likely to end in deadlock, with the weakest models failing to reach any agreement at all under low budgets.
So the two cases to plan for are not clever and stupid. They are funded and squeezed. A funded counterparty will take a firm, well evidenced position without much argument, which rewards having your terms written down and your justification ready. A squeezed one will burn thirty rounds and may walk away with nothing agreed, which rewards deciding in advance how much of your time an unclosed negotiation is allowed to consume before you stop answering.
There is a related number on the repair side. When the researchers tried to improve the weakest model with better prompting rather than a bigger model, out of budget errors fell from 18.4% to 1.3%, while overpayment barely moved and deadlocks did not move at all. Preparation fixes the errors of carelessness. It does not fix the errors of judgement, and that distinction is roughly where a person still has to sit.
One more figure from the same work is worth holding on to, because it quantifies how much the seller side can actually gain. Measured on relative profit, the best performing seller model in the study returned 13.3 times the baseline model's result, with a second model close behind at 12 times. Those are simulation numbers against a deliberately weak reference, so they are not a promise about your margin. They do say that the spread between a prepared seller and an unprepared one is wide enough to be worth an afternoon.
What this does not change
It does not change what your product is worth, and it does not rescue a business that is competing only on price. An agent negotiating on behalf of a buyer is a procurement efficiency, not a new market. If your margin was already thin enough that 3% decides whether an account is viable, the problem predates anything in this article.
It also does not remove the relationship. The published pilot covered smaller contracts with equipment suppliers precisely because those relationships were transactional already. The work that is genuinely relational, a spec that needs explaining or a problem that needs solving, is still done by people, and the organisations deploying these tools say so themselves when they describe freeing their buyers for the large deals.
And it does not make the floor optional. Every failure mode in the research reduces to an agent that had no hard constraint and so found a number by conversation. Set the number first. Let the software find the words.