Agentic Checkout Is a Control Problem, Not a Checkout Problem

Shoppers are willing to let an AI agent buy things on their behalf, as long as a few ground rules are in place, and the rules they care about most have to do with money. 

In our State of Checkout 2026 survey, conducted by Talker Research, we asked 2,000 US adults what they’d need before handing over the shopping. 43% said they’d insist on setting how much the agent can spend, and 41% said they’d want to pick the payment method themselves, even if the AI kept saving them money along the way. 

Both of those rules end up being enforced in the payment credential the agent carries to checkout, so it seems like making agentic commerce work is really a payments problem before it’s anything else. Let’s look at what we found out from consumers for the Report. 

The rules shoppers want their AI agents to follow

Agentic checkout is when an AI agent pays for something on a shopper’s behalf, inside limits the shopper has set. That agent could be in ChatGPT, in a retailer’s own app, or somewhere else entirely, so the limits need to live in the payment credential, since that’s the one piece every agent has to use when it pays.

We asked people which parts of a purchase they’d want to keep for themselves, and how much to spend and how to pay were at the top of the list, at 43% and 41%. Timing came next at 38%, then brand at 35% and retailer at 33%. Only 19% would let AI make all five calls, so most people are picturing an agent that shops for them but still plays by their rules.

What stands out about that list is how practical it is. When people thought about what they’d keep for themselves, they went straight to the decisions that cost them something, and the top two come down to how much money goes out and which card or account it comes from. That’s basically a requirements document for agentic checkout, and since spend limits and payment methods are handled in the payment credential, it’s the payments team that ends up needing to make that happen for consumers.

The ready market is the approval-built-in market

If you only quote one number from this survey, it’ll probably be this one: 36% of US adults would let an AI assistant spend without checking with them first, as long as it kept finding the best price. That 36% is really two groups, though: 20% who said yes, full stop, and 16% who said yes, but only for small stuff. The other 64% said no.

A better way to size the market is to look at how people feel about handing over their payment info. When we asked if they’d give an AI assistant access to it in exchange for the lowest price online, 19% said yes and another 23% said maybe, which means a little over 40% of shoppers are at least open to the idea. The maybe group is the one to watch, since they’re interested and just need a good reason to say yes.

We also asked what would stop people from letting AI buy for them, and the answers line up with that. Fear of fraud came first at 52%, then not trusting AI companies at 45%, then worrying about overspending at 40%. When we asked who they’d blame if an AI-picked purchase went wrong, 47% pointed at the AI company, and 27% weren’t sure. So what people want before they’ll say yes is protection from fraud, a cap on what the agent can spend, and a clear answer about who’s responsible when something goes sideways.

All of that points the same way. The kind of agentic commerce that’s going to win over the next few years is the kind where AI does the legwork and the shopper still gets the final say.

For merchants, that’s good news. Around 40% of shoppers are already open to it, and the way in is mostly a set of controls your payments team can scope, build, and measure.

Payments enforce the limits, and shoppers need to see it happen

In agentic checkout, the limits shoppers care about most get enforced in the payment credential and the authorization that follows it, so they hold up no matter which agent is doing the shopping. That’s only half the job, though. Our report found that the biggest thing missing right now is visibility, which means shoppers also need to be able to see those limits doing their job. The table below maps each thing shoppers asked for to where it’s enforced and what they need to see.

Spend and payment method, the two things shoppers care about most, are already built into the credential by the agent payment protocols. 

OpenAI’s Delegated Payment Spec makes every payment token single-use and caps it at a maximum amount and an expiry time. Google’s Agent Payments Protocol has something called a Payment Mandate, which records the shopper’s limits, including their budget and which payment instruments the agent can use, and passes it along to the credential provider, the card networks, and the merchant’s processor.

Both companies are building agentic products of their own, so their specs tell you where the controls go and say a lot less about how fast anyone will adopt them. Still, it’s worth noticing that two competing protocols handled the two things shoppers ranked highest in the same way.

A card network landed in the same spot. When Visa announced Intelligent Commerce Connect on April 8, 2026, it listed spend controls right alongside payment initiation, tokenization, and authentication as part of its acceptance platform. It also called the product token vault agnostic, which means agent platforms can work with whichever vault they already use. Visa has its own reasons for saying all this, of course. The useful part is where it chose to put spend controls, which is in the infrastructure that accepts payments.

Fraud and accountability are the things that shoppers balk at the most. Fear of fraud is the biggest reason people hold back on letting AI shop for them, at 52%, and 30% point to a lack of transparency. The report’s take is that the industry already knows how to fight fraud, after decades of building fraud detection, tokenization, and dispute handling. What shoppers can’t do yet is watch any of it happen, and AP2 is built with that in mind, with an audit trail that follows every transaction.

When we asked who they’d blame if an AI purchase went wrong, 41% said the AI company first, 6% said the payment provider, and 31% weren’t sure. That 31% is up for grabs, and whoever makes it easy to see who’s responsible gets to claim it. OpenAI’s production guide for merchants already puts refunds and chargebacks on the merchant, so the merchant is on the hook either way and might as well keep the record that shows what happened. The report argues that the payment layer needs this kind of accountability on its own, separate from whatever the AI company promises.

We left brand and retailer out of the table on purpose. At 35% and 33%, they’re decisions about what to buy and where, and those belong to the agent experience. That’s the commerce side of the split we covered in The Difference Between Agentic Commerce and Agentic Payments. Payments can’t pick you a better pair of running shoes. What they can do is make sure the agent spends only what the shopper allowed, on the card the shopper picked.

Your checkout controls stay home when the agent isn’t yours

When a product team builds an AI assistant for its own site, it usually treats spend limits and approval prompts as interface features, so it designs a slider, a confirmation screen, and maybe an animation to go with them.

That works until the purchase starts in ChatGPT, Gemini, or another retailer’s assistant, where none of that interface shows up. A spend cap that only lives in your front end protects your front end, so when a shopper buys through an agent you don’t own, nothing stops that agent from going past it.

Shoppers still need to see their limits, like we covered earlier, but you can’t count on another company’s agent to show them the way you would. So the limits have to sit in the payment layer, which every agent goes through before any money moves.

Our white paper, Future-Proof by Design, says agentic commerce “changes who or what starts the transaction, while the core requirements of execution remain intact.” Two of those requirements matter here. Your systems have to enforce authorization limits automatically, in code, and you have to control how credentials get used without tying that control to any one processor.

Follow that through and you’re left with one place for the controls to live. Someone else owns the agent, and your processor can change by region, by cost, or during an outage. The layer you can control across every agent is a vault and orchestration layer that sits underneath all of them and isn’t tied to a single processor.

That’s the layer Spreedly Agentic is built on, using our Vault and Connect. If you want to see how the agent protocols compete above that layer, read ACP vs UCP: Don’t Bet Your Payment Stack on One AI Agent.

Your 2027 plan needs answers on enforcement, custody, and visibility

If agentic payments goes into the 2027 plan, it belongs there as a governance line item, not a front-end feature. These three questions define the scope, and each has a recognizable good answer.

1. Who enforces the spend limit when the agent isn't yours? A good answer names a control that runs on the credential or authorization path, independent of the agent surface.
2. Where does the credential live, and who can move it? A good answer is a vault the merchant controls, portable across processors.
3. Can you see every agent-initiated transaction across all your processors? A good answer is one log, not one per PSP.

A team that can answer all three has a budget line. A team that cannot has found its scope.

Get the full consumer data

The consumer figures in this post come from the State of Checkout 2026 report, which we commissioned from Talker Research. The full report includes the complete control rankings, the readiness segments, the liability data, and the survey methodology.

‍