AI shopping assistants can be fooled by the same pricing tricks that work on people, but only when nobody tells them exactly what to look for.
Researchers ran eight commercially available large language models from three different companies through a test rig called Tool-Lab, which hides a product's specs behind costly tool calls, mimicking the extra clicks a real shopper has to make to find the fine print. The team wanted to see if classic marketing moves, like pricing something at $9.99 instead of $10, or slapping on a promotional label, could steer an AI agent's purchase the way they steer a human's. When gathering more product information was free, the pricing tricks barely worked. But once digging for details cost the AI extra tool calls and its instructions were vague, it started skipping the numbers needed to work out true cost per item, landing on the same kind of suboptimal picks humans make when they're in a hurry.
That matters because more people are handing routine purchases, groceries, gifts, subscriptions, to AI agents without thinking hard about how they phrase the request. The failure here isn't a hidden flaw baked into the models. It's a product of two things a person actually controls: how specific the shopping instructions are, and how a storefront chooses to structure its product information.
A vague "find me something good" is apparently just as risky a brief for a chatbot as it is for a distracted human standing in aisle nine.