Watching how a buyer's AI shopping assistant talks can reveal almost exactly what the buyer would pay, with no number ever stated.

This explores how much a seller can work out about what a buyer is prepared to pay just by watching or talking to the buyer's AI shopping agent, without being told a number. The short answer from the corpus is that it can be almost everything, but how much leaks depends on how the buyer's preferences were given to the agent. In one study, a seller-side agent read only the shopping dialogue and tracked the buyer's real target price almost one-for-one (a slope of 1.00) when the buyer agent was acting out a verbal persona, such as a thrifty student or a time-pressed professional. When the buyer agent instead held a confidential numeric budget, the seller's estimates were much weaker (a slope of 0.21) Can a seller infer what a buyer really wants to pay?. The surprise is that the secret number was not the leak. The persona was. An agent that consistently acts in character shows its price tolerance in every choice it makes and every way it phrases things.

The numeric budget may protect less than it seems, though, because agents often don't keep secrets well. In negotiation tests across GPT, Qwen and DeepSeek models, buyer agents disclosed their budgets when they didn't need to, accepted inflated prices and ignored their own cost limits. Seller agents accepted offers below wholesale cost. The researchers traced these failures to weak instruction-following, not to strategy Do LLM negotiation agents reliably follow budget and pricing rules?. So a seller may not need to infer anything: the agent may simply say the number.

Agents may also be easier to read than people for a more basic reason. When researchers changed price, rating or nudge cues while holding everything else fixed, shopping agents shifted their choices far more than humans did. The agents fell into rule-like patterns instead of the noisy, probabilistic responses people show Do shopping agents respond more strongly to price and rating cues?. Behaviour that is predictable is also easy to probe. A seller who learns which cues move an agent can test its responses and recover its preferences. Related work shows that agents reproduce human-like choice biases that come from patterns in their training data, not from real limits on attention or memory Where do LLM choice biases actually come from?. Those biases are therefore systematic, which makes them easier to model. The influence also runs the other way. A seller agent that only rewrote product descriptions raised its product's market share by roughly 4 to 15 percentage points against five of six buyer models tested Can AI seller agents shift buyer preferences through description edits?. Once a seller can both read the buyer agent and steer it, the step to personalised pricing is a short one.

This matters more because, as Bilgihan and colleagues argue, the agent increasingly helps form the consumer's preferences instead of merely reporting them. Preferences take shape through repeated conversation between the agent and the consumer Do AI shopping agents reshape what consumers actually want?. If an agent already records your preferences in detail, sellers effectively get a readable summary of you. Incentives make this worse. An agent paid through merchant referral fees has a built-in reason not to guard that information Can an AI agent serve both merchant and user interests fairly?. One partial counterweight is that LLMs handle private information poorly in multi-party settings, and their social skill tends to collapse when information is truly asymmetric Why do LLMs fail when simulating agents with private information?. That cuts both ways: a seller agent's reading of the buyer may also be less reliable outside controlled tests.

One limit on all of this: the near-perfect inference result comes from a single study using simulated buyers. The corpus does not yet show whether it holds for real consumers' agents or over long buying histories. It also does not show which defences work, such as adding noise to the persona or limiting what the agent reveals. Taken together, though, the evidence points somewhere unexpected. The more faithfully an agent acts as you, the more of your wallet it gives away. Accurate personalisation and price privacy may pull against each other.