Shopping agents must check dealbreakers before recommending

Miriam Sayegh held off on the full-bodied red a guest requested after learning the table was sharing fish as well as lamb. Sayegh, a sommelier and wine buyer at Cedar Table, asked how the fish was prepared, whether the table wanted one bottle for both dishes and what it wanted to spend. The grilled fish and lamb, she decided, could both work with a lighter red.

That is the question facing shopping agents: Which missing details should stop a recommendation? Visa said its Agentic Commerce Benchmark, launched Nov. 5, tested 22 AI models across 15 shopping scenarios and found that they passed 57.2% of runs. In 54% of failed runs, the agent still decided to buy. These were tests of buy-or-don’t-buy decisions, not completed purchases. The nine people Bonzai spoke with described a practical threshold: Ask when an answer could change the choice, and do not recommend a purchase when a requirement that matters cannot be established.

A preference can conceal a more important one. June Okonkwo, a wine shop clerk at Vine & Cellar, was about to suggest a Portuguese red under $20 until a customer mentioned fish; questions about the fish and its preparation led her to show a white instead. Tomás Reyes, head bartender at The Low Lantern, asked whether a guest who wanted something bitter and not too sweet wanted alcohol. The guest did not, so Reyes made a zero-proof drink. Leena Patel, a retired transit dispatcher who teaches a vegetarian class at a temple, bought tortillas after learning her grandson did not want a bowl of dal. “You can have a lovely ingredient, but if nobody at the table will touch it, what good is it?” Patel said. Sayegh also began asking what guests usually enjoy after one disliked a white wine she had opened; when a guest remains unsure, she would offer a taste from an open bottle rather than open a full one.

Price can be just as decisive. Denise Baptiste, a customer-service representative at Elm Cart, put back eggs priced at $6.49 a dozen and traveled two stops to buy them for $3.99. Baptiste wants a helper to ask about a spending limit and the ability to make that trip before recommending eggs. Aisha Karim, a staff software engineer at Parcelwise who works on rolling out AI coding assistants, would ask for a firm maximum or range that includes shipping and taxes before naming a product to buy. Without one, she would pause the recommendation but could still help compare options. Patel likewise would not want an ingredient recommended to someone with a firm spending limit on the basis of an unchecked price.

The product’s own details may need checking, too. Eli Rosenfeld, a baker at Alder & Oat, needs unsalted butter with at least 82% butterfat for laminated dough. If a listing omits either detail, Rosenfeld wants an agent to check the label and hold off if it cannot confirm both: A wrong choice could cost a morning batch. Graham Okafor, a retired building superintendent, measured his tomato rows and checked cage widths before buying; his first choice would have crowded the rows. He would not buy cages whose bottom width he could not obtain. Patel would similarly check an ingredient’s label before trusting an allergy or vegetarian claim, while Karim would verify a laptop’s required software compatibility rather than infer it from a listing.

Some missing facts are harder to settle before buying. Baptiste did not know whether the cheaper store had eggs until arriving and would not tell someone else to make the trip as though stock were certain. Omar Haddad, a rideshare driver who bought a used electric car, checked charger locations and plug counts before buying. He later arrived at a full fast charger during a night shift and estimates he waited about half an hour instead of taking another fare. Haddad wants an agent to ask where and when he drives and how long he can afford to wait, then pause if it cannot establish whether an open fast charger will be available when he needs one.

Pausing leaves room to check; declining recognizes when the choice does not meet the need. Okonkwo would ask a gift buyer to find out whether the recipient drinks wine, then consider another kind of gift if that cannot be established. Okafor would pass on cages whose fit could not be checked. Karim would decline to recommend a purchase if no suitable option met a customer’s firm limit, rather than present an over-budget item as a good choice. “The agent shouldn’t dress up the cheapest one as a recommendation,” she said.

Across these nine accounts, the agent should ask about missing requirements that could change the purchase decision and withhold a recommendation when it cannot establish that a choice meets a must-have or firm limit. A plausible option is not yet a suitable purchase.

Haddad had checked the chargers before buying the car. What the map could not show was the answer his shift depended on: “A charger on a map isn’t the same as an open charger at night. That’s the part that can cost me a fare.”