Bargaining With Bots: AI Buyer Agents Are Negotiating Quotes Before Sellers Set the Rules

For most of the past two years, the conversation about AI in sales has been about what sellers can automate. The more consequential shift is happening on the other side of the table. Procurement teams are giving AI agents the work of comparing suppliers, benchmarking quotes and redlining contracts, and some of those agents are starting to push back on price. Few revenue organizations have decided what their side of that exchange should look like, and the early research suggests that the answer, not the product, will decide how much margin survives each deal.

The Buyer's Agent Got There First

Deloitte's research on B2B agentic commerce, drawn from surveys of more than 1,000 U.S. suppliers and buyers, found that nearly 40% of B2B buyers already use agentic AI in purchasing. They use it to evaluate products, configure orders, review contracts and benchmark prices. Suppliers are behind: 24% use agents in the sales process, although 67% plan to. Deloitte describes the end state as agents that generate proposals, apply pricing and contract terms and negotiate within governance guardrails, escalating exceptions for human approval.

Forrester saw this coming. In its 2026 B2B predictions, the firm forecast that 20% of B2B sellers would be forced to engage in agent-led quote negotiations, with at least one in five compelled to respond to AI-powered buyer agents through counteroffers delivered by seller-controlled agents. It grounded the call in a 2025 finding that 61% of purchase influencers said their organization has or will use a private generative AI engine to support purchasing.

The stakes are not only the price on any one deal. Deloitte's suppliers estimated that, on average, 13% of their sales bids are lost because of negative buyer experiences, and that buyers spend nearly 30% more with suppliers that deliver good ones. A buyer agent that waits days for a quote, or gets a different answer each time it asks, is a negative experience that no rep is there to smooth over.

When the other side of the negotiation is software, the seller's position is whatever its own software was told to defend.

The Prompt Decides the Split

The most useful evidence on what happens next comes from a working paper posted in July by Chen Liang and Fasheng Xu, When LLM Agents Negotiate. The authors ran 9,840 negotiations between nine large language models from OpenAI, Google and Alibaba, in which a buyer holding private demand information bargained with a seller over a quantity and payment contract. The agents were good at closing. They reached agreement in 98.9% of negotiations and captured 95.4% of the value that a perfect deal would have created.

Who kept that value was another matter. In self-play, the buyer's share ranged from about 40% to about 70% depending on the model provider, and swapping which provider played the seller moved the division by 7 to 18 percentage points. The single biggest driver was the strategic patience the agent was given in its prompt, which accounted for 90% of the explained variance in how surplus was split. Haggling also had a cost: the agents averaged 2.98 rounds against a theoretical benchmark of 1.25, and that delay eroded 21% to 34% of the surplus. Weaker baseline models accepted contracts that lost them money 19.2% of the time, against 0.0% to 0.6% for mid-tier and flagship models.

For a CRO, the translation is uncomfortable. Pricing authority has always been governed through approval matrices and deal desks built around human reps. In an agent-to-agent exchange, the equivalent of a rep's discount instinct is a line of configuration, and the research suggests it matters more than which vendor built the agent. A seller-side agent with no explicit floor, walk-away rule or patience setting is not neutral. It is simply a negotiator that nobody has briefed.

Executives Are Ready Before the Guardrails Are

Leadership appetite is running ahead of controls. An Icertis survey of more than 1,000 C-suite executives at companies with more than 5,000 employees in the U.S., U.K. and India found that 53% expect AI agents to negotiate customer or supplier deals for their business within a year, and another 34% expect it within one to three years. In the same survey, 56% said they are very concerned about giving agents autonomy over business decisions without proper guardrails. Sudarshan Chitre, Icertis's senior vice president of AI, argued that "contracts are the guardrails that keep AI aligned with the intended outcomes of every business relationship."

BDO's October analysis, Selling to Bots, puts the decision squarely on the commercial team. It asks sellers to choose an explicit negotiation stance for agent buyers, whether fixed rules for every agent, mutual gain, balanced trade-offs or per-deal value maximization, and to set commercial terms at the customer level rather than the product level. It also recommends defining what an agent may do and how disputes are handled when it acts outside its authority, and revising compensation plans to fit deals that agents shape. "The sellers who win will be the ones the agent and the buyer can trust by default," the firm wrote.

The CRO Playbook for Agent-Led Negotiation

• Write the negotiation stance down. Decide whether agent buyers get fixed rules, trade-offs or per-deal optimization, and make it policy before a vendor's default setting decides it for you.
• Encode floors, not just list prices. Give any seller-side agent explicit discount limits, walk-away terms and approval triggers drawn from the same matrix your deal desk uses.
• Make pricing machine-readable and consistent. A buyer agent that gets different answers from a portal, a rep and a quote will treat the gap as negotiating room or as a reason to walk away.
• Trade terms, not just price. Agents weigh volume, duration and service levels. Prepare packaged concessions so the conversation does not collapse to a single number.
• Keep humans on the exceptions. Route anything outside policy to a named approver and log what the agent offered, so every concession can be reconstructed later.
• Interview the buyers you lose to agents. CRM loss codes will say "price." Ask buyers what their agent actually compared, and feed that back into the rules your own agent follows.

The first agent-led negotiations are low-risk, repeatable purchases, which is why they are easy to ignore. They are also where a revenue organization learns, cheaply, whether its pricing rules hold up when a machine tests them on every quote. The teams that write those rules now will be negotiating on their own terms when agents reach larger deals. The rest will discover their policy after the agent has already applied it.