Visa tested 22 AI shopping agents by asking each to buy three common products: a television, a laptop, and a hair appliance. Across the study, the agents succeeded roughly 57 percent of the time. Each agent made 15 attempts, and the results are published on a new Visa website. Agents built on newer models performed best, with a Google Gemini version and a Meta model each succeeding 82 percent of the time, while a Claude agent reached 76 percent. Visa reported that the five priciest models averaged 67 percent success, versus 50 percent for the five cheapest. Visa said failures often involved wrong purchases or no suitable suggestion, since shoppers do not always state every need. Payments and legal liability for rogue purchases remain unresolved.
Why it matters: The results show AI shopping agents still struggle to match consumer needs, a limit that matters for crypto readers tracking how agentic commerce and payments develop.
Summary by Dextape, based on reporting by Fortune Crypto. Written with AI from the publisher’s text and checked automatically for copied wording and unsupported figures. Not investment advice.
Headline and excerpt from Fortune Crypto’s public feed. Fortune Crypto holds the copyright and publishes the full story; Dextape links to it rather than reprinting it.
The Dextape Desk signs Dextape's own work: briefs on wire stories, the daily market report and desk analysis. Briefs are written by an AI model from the publisher's article text and checked automatically for copied wording, unsupported figures and investment advice before they appear. AI drafts from the editorial desk need corroborating sources and an editor's review before release.
