Agentic Checkout Is Being Built Faster Than Anyone Trusts It
Seven percent. That is the share of shoppers who said they would let an AI agent buy something on their behalf, even under conditions they set themselves, in consumer research from ACI Worldwide. More than half said they were flatly uncomfortable with the idea. Those figures sit in the middle of a Fortune report about something moving quickly in the opposite direction: Mastercard and Visa each building the infrastructure that would let AI agents pay for things without a person clicking the button.
The two halves of that report describe the same market from opposite ends. On one side, two of the largest card networks are competing to define how agent-initiated purchases will work. On the other, the people those purchases are meant to serve mostly don’t want them yet. The distance between those two positions is the most interesting thing in payments to me right now, and I don’t think it gets closed by better infrastructure alone.
What Mastercard and Visa are building
The facts first, as Fortune reported them. Mastercard has launched Agent Pay, a virtual card product designed for checkout by an AI agent. The card carries spending limits and retailer restrictions, and the purchase happens without a manual click to pay. Visa is pushing a competing effort called Intelligent Commerce, and the race between the two is about who defines the standards that agentic checkout gets built on. Mastercard also has a mechanism it calls Verifiable Intent: a record intended for resolving disputes over purchases an agent made.
In plain terms, the agent never gets your card. It gets a constrained one, issued for a narrow purpose, with a ceiling on what it can spend and a list of where it can spend it. The limits are what make it reasonable to hand a payment credential to software at all.
I’ll be flat about the standards race itself, because I find this part of the industry tedious. Two networks each want their programme to be the default, both will announce partners, and for a while merchants will be told they need to pick a side. Coverage of contests like this tends to focus on who is ahead. For anyone who runs payments for a living, who is ahead matters much less than what the eventual rules say about who carries the loss when something goes wrong.
Why the 7% matters more than the launches
The standard reading of a number like this is that it’s an early-adoption problem. The argument goes that consumers were once nervous about entering card details online, then about paying with their phones, and they came around each time. Build the infrastructure, make it convenient, and trust will follow.
I think that comparison is wrong in a specific way. Entering a card on a website, or tapping a phone at a terminal, still leaves the person making the decision. Every one of those earlier shifts changed how a payment was made. Agent checkout changes who decides to make it. A shopper who says no to that isn’t being slow to adopt a new payment method. They are declining to delegate a spending decision to something they have no way to check.
That’s a reasonable position, and it’s one I hold in my own work. The AI system I built to run my own sales pipeline never sends anything on its own. Every message it drafts waits for my review, and nothing enters the record without a source it can point at. I didn’t build it that way because the drafts are bad; most of them are fine. I built it that way because an agent reports a finished job in the same tone whether the job was done correctly or not, which I’ve written about in an agent’s self-report is not evidence. The majority of shoppers in the ACI research appear to be applying the same instinct to their own money, without needing an article to arrive at it.
Put plainly, the trust problem in agentic checkout is that a shopper cannot verify the agent’s decision before the money moves. Better checkout flows don’t touch that problem, which is why I don’t expect the launches alone to move the number much.
The dispute record is where this gets decided
Of everything in the Fortune piece, Verifiable Intent is the item I’d watch most closely, because it addresses what happens after something goes wrong. Spending limits and retailer restrictions reduce how much can go wrong in the first place. A dispute record is what determines who pays for it when it does.
Card payments have always come down to one question: did the person whose money this is authorise this transaction. Most of the machinery around them — authentication at checkout, and the dispute and chargeback process afterwards — exists to answer that question, either before the payment or after it. A chargeback is the cardholder’s bank reversing a charge at the cardholder’s request, after which the merchant either proves the charge was legitimate or absorbs the loss.
Agent checkout splits that single question into three. Did the customer give the agent a clear mandate. Did the agent interpret the mandate correctly. Did the agent execute what it interpreted. Each step can fail independently of the other two, and a customer disputing the result has no reason to care which one failed.
Let’s imagine a situation to make that concrete. A shopper tells an agent to reorder a household product they buy every month, sets a spending limit comfortably above the usual price, and restricts it to a few retailers they trust. The usual item is out of stock at all of them. The agent, working within every constraint it was given, buys a larger pack of a similar product from one of the permitted retailers, still under the limit. From the card’s point of view nothing unusual happened: the spend was within the ceiling and the merchant was on the list. From the shopper’s point of view, they were charged for something they never asked for.
That dispute has no obvious owner. The merchant sold what was ordered. The agent stayed inside its constraints. The shopper never approved this specific purchase and is right that they didn’t want it. Resolving it depends entirely on whether a record exists showing what the shopper asked for, what the agent understood, and what it did, in a form that every party in the chain can read and agree on. That is the job a mechanism like Verifiable Intent is being built to do. I don’t know yet how it will work in practice, or how liability will end up being allocated, and I’d be wary of anyone who claims to know at this stage. What I am confident about is that this record, much more than the checkout experience, will decide whether the shoppers who currently say no ever change their minds.
The guardrails in Agent Pay deserve some credit here, because they are better than they might sound. Spending limits and merchant restrictions are structural controls. They don’t depend on the agent behaving well; they cap the damage if it doesn’t. That is the right design instinct, and it’s the same one behind keeping a review step in front of anything my own system sends. What neither control does on its own is prove intent afterwards, and proof of intent is the thing a disputing customer, a merchant and a card issuer will all be arguing about.
Trust-but-verify applies to the infrastructure too
The same discipline that makes me distrust an agent’s account of its own work makes me cautious about any single piece of payment infrastructure, including the new ones.
Seven years in payment orchestration have mostly taught me that no single route through the payments system is reliable in the way people want it to be. Every payment provider is one path through the banking system, behaving the way that path happens to behave this month, and a business with one path has nothing to compare it against. I made that argument in detail in there’s no such thing as a bad payment provider. The fix is redundancy and reconciliation: more than one route, and a record of what each route claims happened, checked against a separate record of what did happen.
Agentic checkout is going to produce the same shape of problem at a new layer. There are at least two competing network standards, each likely to come with its own rules for mandates and disputes. A merchant that integrates deeply with one of them before those rules settle has recreated the single-provider situation: a dependency chosen early for good reasons, which later turns out to be a commitment nobody consciously made. The orchestration answer is the one it always is. Keep your own integration independent of any one programme, so that whichever standard wins becomes a routing decision later rather than a rebuild.
Reconciliation carries over as well, and it matters more with agents, not less. In ordinary payments, you check what a provider reports against your own order records. With agent-initiated purchases, your own order records will need to hold something they have never had to hold before: evidence of what the customer instructed, kept separately from what the agent did. A merchant that records only the transaction will have the agent’s version of events and nothing independent to check it against, which is exactly the situation my own system is built to avoid.
Where this argument stops
A survey of stated preference is a snapshot, not a forecast. People regularly say they won’t do something and then do it once it becomes convenient and nothing bad happens to anyone they know. The 7% could look very different in two years, and if it does, the infrastructure being built now is what will carry that volume. I don’t think Mastercard and Visa are wrong to build it. I think they are building the easier half first.
It also doesn’t follow that every merchant needs to act on this now. If you run a single-market business with one provider that covers your customers well, there is no agentic-checkout project for you this year, and building an orchestration layer in anticipation of one would be spending money against a risk you aren’t running. The same applies to most smaller businesses generally. The useful thing to track is how dispute rules for agent purchases develop, and the time to move is when they are written down.
For businesses running real payment infrastructure at volume, across several providers and markets, the practical list is short. Don’t tie your checkout to one network’s agent programme while the standards are still being contested. Start working out what an order record would need to contain to defend a disputed agent purchase, because producing that evidence is going to fall to you. And treat every agent-initiated transaction the way you would treat any other output from an agent: probably fine, not yet verified, and worth keeping a separate record to check it against.