“Help me choose a laptop for work. Nothing too expensive.”
Previously, you might have compared specifications across several websites and asked a friend for advice. Now you can give the request to AI and ask it to shortlist options, compare prices and even prepare a purchase.
But when all you have said is “for work”, does the AI really know what you need?
On 5 November, Visa published an AI shopping benchmark involving 22 models from 12 labs across 15 shopping scenarios. The overall pass rate was 57.2%, meaning 42.8% of test runs failed. The best-performing models passed 82.2% of attempts.
The figures need context. The benchmark tests whether AI can uncover a customer's requirements through conversation and correctly decide whether to buy. It does not measure payment failure rates in the real market. Visa plans to extend testing to checkout, payment and order placement. In 54% of failed runs, the AI still chose to buy, despite missing requirements or selecting an unsuitable product.
The hardest part of shopping may be establishing what the customer actually needs.
For a laptop, “for work” can mean many things. Someone carrying it between appointments may prioritise weight and battery life. Another person needs particular accounting software. Someone else mainly answers emails but must connect two external monitors at the office.
If AI recommends a product based on reviews, price and popularity before asking those questions, its answer can look comprehensive and still lead to the wrong purchase.
This is a hypothetical scenario, but a familiar one for businesses. An experienced salesperson usually asks follow-up questions to understand the gap between the customer's opening request and the factors that will actually determine their decision.
Businesses designing AI shopping assistants would benefit from making that discovery step part of the workflow.
Suppose a customer wants a printer. The system should first establish monthly printing volume, whether colour is needed and whether double-sided scanning is required. It can then compare suitable models. If none of the available products fits, it should be able to say clearly: “There is no suitable option at the moment.”
That response may not produce an immediate order, but it can reduce mismatched purchases, returns and subsequent complaints.
Merchants preparing to introduce AI can test situations where the customer leaves out information, their budget conflicts with their requirements, or every stocked product is unsuitable. Watch whether the system asks more questions, acknowledges its limits and knows when to hand over to a person.
One successful recommendation shows that the system succeeded on that occasion. Consistently serving different customers provides a stronger foundation for deployment.
Payment permissions can be introduced gradually according to the task. Start with research and comparison, then preparation of a shopping list. Before an order is placed, show the model, quantity, total cost and any unresolved conditions so the user can make the final decision.
An AI shopping assistant earns trust through repeated demonstrations that it understands correctly, makes reasonable choices and pauses when it lacks enough information.
When AI tells you “this is the best option for you”, there is another question worth asking: how much did it actually establish about your needs?
Source: Visa | AI shopping benchmark (5 November 2026). This article includes independent editorial analysis by SRAM.
Author: SRAM