Only 23% of U.S. consumers surveyed for Visa said they trusted generative AI to handle payment transactions on their behalf, exposing a substantial gap between using AI to shop and allowing software to spend money.
The finding matters because payment networks, AI platforms, banks and merchants are moving agents closer to checkout. Visa says more than 100 companies are integrating agents with its Visa Intelligent Commerce network and more than 150 issuers are testing agent-initiated payments. Those are company-reported participation figures, not evidence that autonomous payment behavior is already widespread.
The accountability question is no longer simply whether an agent can complete a transaction. It is whether every participant can prove who authorized it, what limits applied, which product and seller the consumer approved, and how an error or unauthorized purchase will be reversed.
Shopping use has outpaced payment trust
Visa said 72% of U.S. consumers had used an AI assistant as well as search engines to discover products. Independent trade publication Digital Transactions described the result as 72% having used an AI assistant for shopping, while only 23% trusted AI to manage payments.
The Visa Trust Index was conducted on Visa’s behalf by The Harris Poll through its Omnibus survey platform. It was fielded in the United States from May 26 through May 28, 2026, among 2,065 consumers matched to the U.S. general adult population using Census benchmarks. Visa said the payment question used samples of 1,028 to 1,034 respondents for each brand tested.
Visa also reported that trust increased to 61% when respondents were asked about a transaction secured by Visa. That brand-conditioned result should be read narrowly. It measures stated trust in a survey; it does not demonstrate that branding, network participation or any specific control reduces unauthorized transactions, mistaken purchases or disputes in production.
The public methodology on Visa’s article identifies the field dates, survey provider and sample sizes, but does not publish the full questionnaire, response options, weighting detail, margin of sampling error or complete brand-level cross-tabs. Those omissions limit how precisely readers can compare the 23% and 61% figures or generalize them to particular payment products and use cases.
Who must make an agent’s authority verifiable
Visa identifies two linked trust problems. First, merchants have spent years trying to block bots and may struggle to distinguish a legitimate shopping agent from hostile automation. Second, consumers need confidence that an agent will remain within the authority they granted and that they can regain control when something goes wrong.
That divides responsibility across the payment chain. The AI platform must preserve the user’s intent and avoid silently expanding it. A wallet or credential provider must bind a payment instrument to that authority. Merchants and acquirers need reliable signals that distinguish an authorized agent from fraud. Issuers still have to decide how to authenticate, approve and monitor unfamiliar transaction patterns. Networks can carry credentials and agent-related signals, but their presence does not remove the duties of the other parties.
Visa says it has deployed tokenized credentials designed for agentic commerce so agents can transact without receiving a consumer’s underlying card number. Tokenization can reduce exposure of account credentials, but it does not by itself answer whether the consumer authorized the exact purchase, whether an agent selected an unsuitable seller, or who bears the loss after a mistake.
Controls must cover intent, not just identity
For payments professionals, a durable control model should capture the scope of the consumer’s mandate before authorization. That may include a maximum amount, approved merchant or merchant category, product constraints, delivery terms, expiration time and whether the agent can substitute items or accept a changed price.
The transaction record should also preserve evidence linking the consumer, agent, merchant, credential and final order. If the payment message shows only that a valid token was used, customer-service and dispute teams may be unable to determine whether the agent followed its instructions.
Human confirmation remains important for higher-risk or ambiguous decisions. An agent that encounters a new merchant, a material price increase, recurring billing, restricted goods or an unexpected delivery address should not treat a general shopping instruction as unlimited payment consent. Clear escalation rules can keep convenience from becoming authority by default.
Operational readiness extends beyond authorization. Merchants, issuers, networks and AI platforms need defined procedures for cancellations, returns, chargebacks, duplicate orders and complaints. Consumers should be able to see that an agent initiated a transaction, review the mandate it relied on and revoke future authority without navigating multiple providers.
Visa’s survey does not show that these controls have failed in a documented incident. It does show that most respondents were not yet willing to hand payment execution to generative AI. For an industry racing to enable agentic checkout, that is a warning against treating technical capability or familiar branding as a substitute for explicit, auditable consent.