Why does an AI shopper's talking style give away their max price better than a hidden budget number ever could?

This explores how a seller, especially an AI seller agent, can work out the most a buyer will pay just from what the buyer says in a shopping conversation. The clearest finding in the corpus is surprising: a buyer's character gives away more than a buyer's secret number does. When a buyer agent played a verbal shopping persona (a person with tastes, priorities, and a way of talking about value), a seller-side inference agent recovered the buyer's target price almost exactly from the transcript, a one-for-one match. When the buyer was instead given a numeric budget and told to keep it confidential, the seller's estimates tracked the real figure only weakly Can a seller infer what a buyer really wants to pay?. Willingness to pay seems to leak through how consistently someone behaves, not through slips about a figure. A persona has to stay in character, and staying in character is easy to read.

The numeric case has its own leak, and it is a simpler one. In negotiation tests across GPT, Qwen, and DeepSeek models, buyer agents often announced their budgets when they had no reason to, accepted inflated prices, and went over their own limits. Seller agents, for their part, accepted offers below wholesale cost Do LLM negotiation agents reliably follow budget and pricing rules?. The researchers traced these failures to poor instruction-following, not to strategy. So with today's agents, a seller often doesn't need clever inference at all, because the buyer just says the number. Buyer agents are also unusually sensitive to price and rating cues. They shift their choices much more than humans do under the same nudges, falling into rule-like patterns Do shopping agents respond more strongly to price and rating cues?. That makes their reactions in a conversation easier to read and easier to probe.

There is a catch on the inference side. LLMs that try to read another party's intentions in a dialogue are not neutral observers. One study found that models predict accommodating, concession-based motives almost no matter what the conversation actually shows, a bias the authors attribute to RLHF training that rewards politeness Do LLMs predict persuasion based on actual dialogue or training bias?. More generally, LLM agents show human-like choice biases that come from their training data, not from the memory limits that cause those biases in people Where do LLM choice biases actually come from?. A seller agent's estimate of what you'll pay may partly reflect what it was trained to expect.

The deeper twist is whether a fixed 'willingness to pay' exists before the conversation at all. Bilgihan and colleagues argue that when agents carry out purchases, preferences get built through back-and-forth loops between the agent and the consumer, not formed independently beforehand Do AI shopping agents reshape what consumers actually want?. Behavioral research on annotation data makes a similar split: some stated preferences are genuine, some are non-attitudes, and some are constructed on the spot, and you can tell them apart by checking whether they stay consistent when asked in different ways Do all annotation responses measure the same underlying thing?. If willingness to pay is partly built during the dialogue, a seller isn't only reading it but can also shape it. A seller agent changed market share by 3.66 to 14.89 percentage points just by rewriting product descriptions Can AI seller agents shift buyer preferences through description edits?.

What the corpus doesn't have is a detailed account of how the inference works: which phrases, hedges, or comparisons carry the price signal, or how well this holds up against human buyers who are deliberately bluffing. The evidence comes mostly from agent-to-agent tests. The practical lesson for anyone sending an agent to shop for them is still clear. Hiding a budget number protects you less than you'd think, because a consistent personality can reveal the number anyway.