Visa’s test found shopping agents completed purchases in about 57% of trials, underscoring a challenge for systems expected to turn broad requests into appropriate orders.
Visa evaluated 22 agents, giving each 15 shopping expeditions involving televisions, laptops and hair appliances. Some tasks left shopper preferences unstated, requiring the software to ask follow-up questions before proposing an item. One television request specified a 30-day return period, Dolby Vision and use with a child’s gaming console.
Visa included cases in which an agent’s recommendation did not fit the request or it offered no appropriate choice among the unsuccessful runs. At the top, a pair of agents each recorded 82%; an OpenAI agent placed sixth with a 64% score.
Results differed across price groups. The five highest-cost agents averaged 67% success, while the five lowest-cost agents averaged 50%. The figures show a gap but do not establish its cause.
Visa Global Head of Growth Rubail Birwadker said, “We’re not there yet from a trust or technology perspective.” Visa and Mastercard are developing payment tools and educating merchants about agent-led purchases, while responsibility for unauthorized transactions and fraud involving shopping agents remains unresolved.
Shopify’s model keeps the shopper in control of the final step: a buyer-approved checkout for AI shopping agents requires the customer to accept the total and payment method before an order can proceed. Visa has made the tested agents available for consumers to use through its results site.
