Consumers Will Let AI Shop. Not Pay Yet.

Financial institutions have spent the past year racing to give artificial intelligence a seat at the checkout counter. New research from Visa suggests the industry is ahead of its customers: shoppers have embraced AI assistants for browsing and comparing, but most still refuse to let one complete a purchase on its own.

The Adoption Curve Has Outrun the Trust Curve

Visa’s newly published Trust Index, built on a Harris Poll omnibus survey of 2,065 U.S. consumers fielded May 26 to 28, 2026, found that 72% of respondents have already used an AI assistant, evidence that the technology has moved well past early adopters. But when the question shifts from browsing to paying, the number collapses. Only 23% of U.S. consumers say they trust generative AI to handle a payment transaction on its own, according to the survey (sample of 1,028 to 1,034 respondents per brand tested on the payment question, matched to U.S. Census adult population data).

That gap between using the technology and trusting it with money is the real story in agentic commerce right now, and it is a bigger constraint on the category’s growth than any remaining technical limitation. Retailers and payment processors have built the plumbing for autonomous checkout faster than they have built the case for handing an agent a card number. Part of the gap is a liability question consumers have not yet been given a clear answer to: when a shopping agent buys the wrong item, the wrong quantity or from the wrong merchant, it is not obvious to most people whether the dispute process that already exists for a mistyped human purchase applies in the same way, or applies at all.

Whoever Wins the Handoff Wins the Rail

Visa’s data also points to where that trust settles once consumers are willing to extend it. Sixty-one percent of respondents said they would trust Visa specifically to handle an agentic transaction on their behalf, a figure that climbs to 68% among 18-to-34-year-olds and 71% among frequent AI users, the two cohorts most likely to actually try the feature. Oliver Jenkyn, Group President at Visa, framed the shift as structural rather than incremental: “AI has the potential to fundamentally reshape how people discover, buy and pay for goods and services, much like e-commerce and mobile commerce did before it.”

The framing is self-interested. Visa sells the underlying rails through its Visa Intelligent Commerce program, which the company positions around secure transaction handling, identity verification, authentication and consumer controls for AI-driven shopping specifically, rather than treating agentic checkout as a variant of an ordinary card-not-present transaction. That distinction matters: the underlying pattern lines up with how agentic commerce has already split into reasoning and payment layers. The model that recommends a product and the rail that actually moves the money are becoming distinct businesses, and consumers appear to trust the rail far less than the recommendation engine, and far less than they trust a network that has spent decades building fraud liability and dispute infrastructure around card payments.

What the Trust Gap Means for the Finance Leader

For a bank, card issuer or payment processor building agentic checkout features, the Visa data argues against a bolt-on approach. A chatbot that can complete a purchase is not, on its own, a product consumers want; a purchase flow with visible spending limits, revocable authorization and a familiar brand standing behind the transaction is. The frameworks already emerging for how an AI agent authenticates and authorizes a card payment, such as EMVCo’s ground rules for AI agents paying by card, matter more to adoption right now than the sophistication of the underlying model.

The practical implication is that trust in agentic commerce is being built at the rail level, not the interface level. Institutions that already hold consumer trust in traditional payments, whether a bank brand, an established card network or a regulated wallet, have a transferable advantage over AI labs and retailers trying to build payment trust from nothing. That advantage will not last indefinitely if incumbents move slowly. Visa’s own investment in Intelligent Commerce is a bet that this window closes within a product cycle or two, not several years, and every issuer sitting on an established brand relationship with a large card base is implicitly making the same bet by default, whether or not it has shipped anything agentic yet.

The generational split in the data is also a planning signal, not just a marketing one. A gap of seven points between the general population and 18-to-34-year-olds on trusting Visa specifically for agentic payments suggests the younger cohort is not simply more trusting of AI in the abstract; it is more willing to extend trust to a specific, named financial brand once it demonstrates the guardrails exist. That argues for institutions to be explicit and visible about their controls rather than assuming a younger, more AI-fluent customer base will tolerate a black-box agent by default.

What to Evaluate Now

Institutions should treat the 23% figure as a design constraint rather than a market-sizing problem. The practical questions for anyone shipping agentic payment features this year: does the flow expose a spending cap the consumer sets and can change at any time, can authorization be revoked mid-session without contacting support, and does the consumer see, in plain language, which party (the agent, the merchant or the network) is accountable if the transaction goes wrong. Visa’s research suggests consumers already have an intuitive answer to that last question, and it is rarely the AI itself. Whichever payment brand makes that accountability visible first, rather than assuming trust in the model will simply transfer, is the one most likely to convert the 72% who already use an AI assistant into the smaller group that eventually lets one pay.

Source: Visa