Consumers are ready for AI to shop. They are not ready for it to pay.

Fifty-eight percent of US consumers say AI is useful for researching products and comparing prices. Thirty-nine percent are using it more while shopping than they were three months ago. And 56% say they would never allow AI to spend money without their approval.

One percent want shopping fully automated.

That gap is where the next phase of commerce will be settled. Consumers have accepted AI as a way to think about a purchase. They have not accepted it as a way to complete one. The reasons they give are specific enough to read as a product requirement, and almost all of them sit in the payment layer rather than the model.

The concern is the money, not the model

Asked what worries them about letting AI buy on their behalf, 76% of consumers name fraud or unauthorized charges. The next answer, buying the wrong item, sits at 33%. Nobody seriously doubts the agent can find the right product.

Reversibility is close behind. Eighty-four percent say the ability to cancel or reverse an AI-initiated purchase after checkout is extremely or very important, with 64% choosing extremely important. Consumers are not asking for a smarter agent. They are asking for a charge they authorized, a limit they set, and a way to undo it if something goes wrong. Those are payments questions.

The industry has spent two years competing on model quality and protocol position. The research suggests that competition is close to table stakes. What consumers are waiting for is the layer underneath.

Consumers described the controls before the industry shipped them

Asked what would make them more comfortable, 61% choose spending limits and 59% choose human approval checkpoints. Every other option, including tokenized payment details at 24%, sits far below.

That is close to a specification. In July, Nuvei completed a live agentic commerce proof of concept with Visa, Arvato Systems, and the fashion brand Kings and Priests. A merchant AI agent initiated a purchase on a shopper’s behalf and paid inside the agent, with no hand-off to a separate payment flow. Multiple issuers across Europe authorized it on live Visa rails using a tokenized Visa credential within Visa Intelligent Commerce, governed by shopper-set guardrails including spend caps and approved categories.

Spend caps and approved categories are the same two controls consumers put at the top of the list. The alignment is not a coincidence. Both the merchants on Nuvei’s Global Customer Advisory Board and the consumers in this research arrived at the same conclusion: delegation works when it is bounded and observable.

Know Your Agent, one of the two building blocks of Nuvei Agentic, extends that logic. It registers and credentials agents, validates the consumer’s mandate before money moves, scores agent reputation, and keeps every action auditable. Who is this agent, did the consumer authorize this, and can it be proved later are questions the industry has never had to ask, because there was always a person at the keyboard. The research shows consumers asking them now.

The purchase is still expected to happen on the merchant’s ground

Fifty-eight percent of consumers would still prefer to complete an AI-assisted purchase on the retailer or brand’s own site or app. Twelve percent would complete it inside the AI platform. Sixty-two percent are not comfortable letting AI complete a purchase without a traditional checkout page at all.

For merchants weighing where to invest, that is the number that matters most. First-party agents, the AI experiences a merchant runs on its own site and inside its own customer relationship, are not a staging post on the way to public agents. They are where consumers say they want the transaction to land.

It also lowers the cost of learning. A merchant that builds this channel on ground it owns keeps the customer relationship and the merchant-of-record, and extends something that already works when public agents arrive. The alternative is starting from scratch inside somebody else’s environment, on terms set by whoever owns the model.

The purchase is still expected to happen on the merchant's ground

Fifty-eight percent of consumers would still prefer to complete an AI-assisted purchase on the retailer or brand's own site or app. Twelve percent would complete it inside the AI platform. Sixty-two percent are not comfortable letting AI complete a purchase without a traditional checkout page at all.

That is a finding about where the transaction lands, not about who builds the agent. Whatever discovers the product, consumers expect the payment to complete somewhere they recognize, under the brand they chose to buy from. Merchants that keep the merchant-of-record and the customer relationship at that moment are holding the position consumers say they want them in.

It is also why first-party agents came first in the pilot. Nuvei's proof of concept ran on the merchant's own platform, which is the environment consumers are most comfortable with and the one where a merchant can learn this channel without ceding the relationship. The same controls extend to public, third-party agents as those arrive. The question for a merchant is not which agents to build. It is whether the payment layer can accept a purchase from any of them, on the merchant's own terms.

What this changes for a payments decision

For a head of payments or a chief revenue officer, agentic commerce is an architecture choice being made now on incomplete information. The research points at three things worth protecting.

The first is optionality. Agent standards are still moving. A protocol compatibility layer lets a merchant integrate once and accept payments from whichever standard an agent uses, whether the Agentic Commerce Protocol, AP2, or the Model Context Protocol, routed across networks rather than betting on one. Nuvei intends to certify against both Visa Intelligent Commerce and Mastercard Agent Pay.

The second is control. Mandate validation, spend caps, agent identity, and a complete audit trail are what turn a consumer’s stated conditions into something a merchant can actually operate. They also determine whether a disputed agent purchase can be resolved or simply becomes a chargeback.

The third is not rebuilding. The components an agent payment depends on already run at scale: a Level 1 PCI-certified environment, local acquiring across the globe, local and alternative payment methods, and real-time fraud and risk tooling. Reaching agentic capability through the platform a merchant already uses is a different proposition from re-engineering a payments stack for a channel that is still forming.

The specification is already written

McKinsey projects agentic commerce will drive roughly $1 trillion in global transaction volume by 2030 and $3 to $5 trillion by 2035. Volume at that scale does not arrive because consumers grow enthusiastic about AI. It arrives when delegating a payment stops feeling like a risk.

Consumers have been precise about what that takes. A limit they set. A moment where they can say yes. A way to undo it. Someone identifiable to hold responsible. Every one of those is a payments primitive, and every one of them lives in the execution layer rather than the model.

‍