Visa test: AI shopping agents failed about half their tasks

Summary

• AI models failed roughly half of defined shopping tasks in a Visa test announced in November.
• The tested stage excluded payment and order placement, which Visa says it plans to add later.
• Shoppers get faster shortlists but rarely learn why certain products surfaced and others never did.

The latest

AI models failed in roughly half of their attempts at defined shopping tasks, according to a test Visa announced in November. The published stage of that test stopped short of completing payment and placing the order — two steps the company says it intends to add later. The results land as payment and technology firms build systems letting agents act with user approval.

Details

• The test: Visa said in November that AI models fell short in close to half of the attempts across a set of defined shopping tasks. The company said the announced phase did not cover checkout or order submission, and that both stages are planned for a later round of testing.
• The trust gap: In a Visa survey in the United States, the company said 72% of consumers had used an AI assistant, but only 23% trusted one to carry out payment transactions on their behalf. The gap separates asking a machine for a suggestion from authorizing it to spend.
• Who benefits: Barron's, in a report by journalist Teresa Rivas, examined whether large retailers stand to gain most from the spread of shopping agents, citing their brand presence, inventory breadth and fulfillment capacity as advantages machines are likely to reward.
• The research: The magazine drew on work by UBS analysts who found that shopping tools frequently recommended large retail chains, among them Walmart and Costco. The finding does not establish that the preference holds across every product category and market, or that smaller merchants are automatically excluded.
• The new battleground: For companies, the quality of data exposed to machines may itself become part of the competition for customers: product descriptions, inventory updates, price clarity, and the availability of delivery and returns information. Consumers may receive a short list quickly, but they do not always know which options never surfaced.
• Visa's stake: Visa holds a commercial interest in developing this market, context that matters when reading its findings and initiatives. The company is working with partners on tools to verify an agent's identity and authority.
• The unresolved question: The spread of payment rails does not by itself settle liability when an error or a fraud occurs — who answers for a transaction an agent executed remains open.
• The UAE move: Visa announced in May a program in the United Arab Emirates to support sector readiness for commerce executed by AI agents, in cooperation with financial institutions and technology partners. The announcement did not specify the participating entities.

Background

AI shopping tools let a user describe a need and a budget instead of moving between multiple sites and stores. Their autonomy varies: some only recommend, while technology and payment companies are building systems that let agents take further steps with user approval.

Between the lines

The UBS finding that tools favored Walmart and Costco points to a shift in where retail competition happens. If machines reward inventory breadth and clean data, visibility becomes a function of how readable a merchant is to an algorithm rather than how appealing it is to a browsing human. Smaller stores in markets like the UAE may gain reach, but only if their product and pricing data is accurate and machine-usable.

What's next

Watch for the next phase of Visa's test, which the company says will add checkout and order placement. Also watch whether agents can explain their selections, respect spending limits, and leave the final payment decision with the user.

Source: Barron's, Visa, UBS