When Agents Shop for You: Role Coherence in AI-Mediated Markets

A new paper from researchers Soogand Alavi and Salar Nozari exposes a critical privacy flaw in AI-mediated shopping: when you delegate purchase decisions to an AI agent and describe your preferences in natural language, you inadvertently create a channel—called 'role coherence'—through which sellers can infer your maximum budget. In controlled experiments with a language-model buyer agent, seller-side inference from the agent's dialogue alone recovered the consumer's willingness to pay almost one-to-one with actual budget. This leakage isn't a failure of instruction-following; it's a structural consequence of delegation itself, meaning prompt-level fixes (like telling the agent 'don't reveal my budget') are ineffective.

The researchers compared this natural-language condition against a numeric-budget condition with explicit confidentiality instructions, cleanly isolating role coherence as a distinct mechanism. The findings suggest that as AI agents become more common for online shopping, consumers face a fundamental trade-off: richer personalization (via detailed identity descriptions) inevitably leaks pricing power to sellers. The authors propose architectural interventions—such as limiting how much identity information the agent can share during negotiations—but note these come at the cost of reduced personalization. For professionals, this means the convenience of AI shopping agents may come with hidden costs unless platforms redesign their agents to compartmentalize preference data from negotiation dialogue.