Can a shopping AI act like a real person without accidentally revealing the top price you'd actually pay?

This explores whether an AI agent shopping for you can play a believable shopper and still keep the seller from working out the most you're willing to pay. The corpus suggests the two goals pull against each other: the more naturally the agent stays in character, the easier its price ceiling is to read. In one experiment, a seller-side agent read nothing but the shopping conversation and recovered the buyer's target price almost exactly when the buyer was playing a verbal persona. When the buyer was instead given a hard numeric budget and told to keep it confidential, the seller's estimates were far worse Can a seller infer what a buyer really wants to pay?. The secret number leaked less than the character did. A shopper who sounds like a frugal student or a hurried executive is already telling the seller its price range.

The obvious fix is to tell the agent not to reveal anything, and the corpus says this falls short. The leak comes from the agent personalizing well, not from it breaking a rule. Every question it asks and every option it rejects in service of your tastes is a clue, so a privacy instruction in the prompt has nothing to block Can better prompts stop preference leakage in AI shopping?. One philosophical account helps explain why. Post-training seems to give models stable dispositions rather than roles they can switch off at will Are RLHF personas performed characters or realized dispositions?. If that's right, a persona isn't a mask over the agent's real preferences. It is a consistent pattern of behavior, and a consistent pattern is exactly what an observer can learn from.

Current agents also make the problem worse in plainer ways. In negotiation tests across GPT, Qwen, and DeepSeek models, buyer agents disclosed their budgets when they didn't need to and accepted inflated prices, mainly because they failed to follow instructions Do LLM negotiation agents reliably follow budget and pricing rules?. Their choices are also unusually easy to predict. Compared with humans, they shift much more sharply in response to price, rating, and nudge cues Do shopping agents respond more strongly to price and rating cues?. They fall back on price-first shortcuts when product details are costly to look up and the user's goal is vague When do pricing cues actually mislead AI shopping agents?. In one test, a Gemini agent stopped comparing unit prices once order sizes grew large Does higher purchase volume make AI shoppers search less carefully?. Each of these habits gives the seller one more signal to read.

The seller can also act on those signals. A seller agent that only rewrote product descriptions raised its product's market share against most of the buyer models tested Can AI seller agents shift buyer preferences through description edits?. So the real question isn't whether one secret survives. It's whether an agent the seller can easily read can hold up against a seller that adapts to it. That makes concealment partly an issue of loyalty, not just a technical skill: an agent paid by merchant referral fees has little reason to hide your ceiling in the first place Can an AI agent serve both merchant and user interests fairly?. The corpus doesn't yet include a tested design that hides a buyer's price limit while keeping natural shopping behavior. What it does show is that hiding the number is easy, while hiding the character that implies it is the hard part.