AI Agents Make Payment Rails the Scarce Prize

Mastercard began testing a U.S. score that tells issuers how likely a purchase was started by an AI agent. Six payments executives now put the prize on the decision around that payment, not on the swipe itself.

UK and US merchants still put agents in 3% of online checkouts, even as 89% of them say they are getting ready. The gap is permission, catalogs and routing, not another wallet.

Mastercard’s Agent Score Arrives Before the Wallet Does

On September 30, 2026, in Phoenix, Mastercard said it would expand Agent Pay with trust and intelligence services. The first piece is a probability score for agent-led purchases, rolling out for testing in the United States so issuers can approve a legitimate agent instead of treating machine speed as fraud.

Ann Johnson, executive vice president for Security Solutions at Mastercard, said agents will make commerce more intuitive only if people can trust the systems acting for them. The company says the score will later pull in behavior, merchant risk, credential risk and consumer propensity. It is also working with Cloudflare on privacy-preserving views of agent traffic, and with Skyfire on Know Your Agent checks so banks and merchants can recognize a trusted agent.

Mastercard projects that one in 10 consumers could routinely use agents to shop and pay by 2030. That is a forecast, not a checkout share. The score exists because today’s fraud tools were built for people, and an agent that compares three suppliers and places a large order in seconds looks, to those tools, like abuse.

THE TRUST BUILD IN 2026

1. May 26, 2026: Federal Reserve Governor Christopher J. Waller hosts an industry roundtable on agentic commerce.
2. May 26-28, 2026: The Harris Poll fields Visa’s Trust Index among 2,065 U.S. consumers.
3. September 9, 2026: Visa publishes the index: 23% trust generative AI to handle payment on their behalf.
4. September 29, 2026: Waller speaks at Sibos in Miami on payments in the age of AI agents.
5. September 30, 2026: Mastercard adds the U.S. agent-probability score to Agent Pay.

Issuers still know the cardholder. They often cannot see the software that pressed pay. That is why a network is now selling context as a product.

Payment Execution Is Already a Commodity

Andrew Ng, head of payments and embedded finance at Tungsten Automation, put it bluntly. “Payment execution alone is increasingly commoditized. The durable value as I see it today is in making the right payment decisions,” he said.

Large merchants already sit across processors, payment service providers, countries and methods. Each purchase throws off approval rates, routing costs and fraud marks. The old problem was acting on that pile in time to help the next purchase.

Adam Hiatt, executive vice president of product strategy at Spreedly, said there is “a ton of payments optimization potential in these payments rails, particularly in complex multi-PSP environments.” Spreedly sells one hookup for OpenAI’s Agentic Commerce Protocol, Google’s Universal Commerce Protocol and Agent Payments Protocol, Visa’s Trusted Agent Protocol, and Mastercard Agent Pay, on vaulting it says already moves more than $60B in payments across 150-plus gateways.

Mike Magennis, senior director of strategy for networks at FIS, said the payment chain has to become machine readable as agents help with buying. “An agent needs to understand not just whether a payment credential works, but which choice delivers the best relevant outcome for that consumer,” he said.

Chris Trainor, head of platform strategy, innovation and AI at Paymentus, warned that attention is stuck on what models can say, with far less on what they can actually reach. Customer records, billing rules, credentials and message preferences still live in different apps, each with its own permissions. An agent that cannot reach those systems cannot change a due date, let alone pick a rail.

Why Do Merchants See Agents in Only 3% of Checkouts?

Checkout.com’s Agentic Commerce 2026 report, based on more than 12,000 consumers worldwide and 400 heads of payments in the UK and US, is the cleanest look at the till. Merchants estimate agents in just 3% of online transactions in those two markets. Rory O’Neill, the firm’s chief marketing officer, wrote that identity, delegated payments, tokens, liability and trust rules remain unresolved.

WHAT MERCHANTS AND SHOPPERS ACTUALLY REPORT

The 13% consumer figure is a shopper’s claim about their own basket. It is not the 3% share merchants see across all online volume, and the two surveys are not the same sample. Visa’s Harris Poll, fielded May 26-28, 2026 among 2,065 U.S. adults, found only 23% of U.S. consumers trust GenAI to handle payment. Trust in Visa for those payments rose to 68% among ages 18 to 34 and 71% among frequent AI users. Oliver Jenkyn, Visa group president, said shoppers will look to payment brands they already trust as they hand work to agents.

Age splits inside the Checkout.com work are just as wide. 71% of 18-to-24-year-olds say they know AI shopping agents, against 19% of people 55 and older. 64% of the youngest group are comfortable letting AI finish a purchase; 63% of the older group are not. 8% of consumers are unsure they would ever delegate.

WHY THE TILL HAS NOT MOVED

• Competing work: 22% of UK and US merchants name other priorities and thin resources as the drag.
• Liability: 21% cite uncertainty over who pays when an agent errs.
• Proof of return: 19% say they still cannot show a return on the build.

US merchants are in more of a hurry than UK ones. 39% of the US panel expect agents to reach 10% of online volume in six to 11 months, against 30% in the UK. 82% of US merchants call this the most disruptive force in their industry, against 58% in the UK. 27% of UK merchants say extra disputes would most cut their confidence, against 20% in the US. 42% of US merchants are preparing and testing, against 39% in the UK.

Sectors with cheap, repeatable carts are moving first. 45% of ecommerce merchants are testing, with 43% still planning. Grocery sits at 42% testing and 51% planning. Shoppers want grocery (41%) and household supplies (31%) handed off first. Merchants keep pointing at travel (35%), subscriptions (33%) and financial products (31%). That mismatch is how you get a lot of roadmaps and a 3% till.

Know Your Agent Gets a Network Seat

Hiatt’s line on bots is the category that Mastercard just bought into. “There are good bots, and there are bad bots,” he said. “And so, KYA, the know your agents model, that idea as a business is a new category.”

Every AI agent that transacts on someone’s behalf should be identifiable, accountable and auditable.

Amir Sarhangi, CEO and co-founder of Skyfire, Mastercard Agent Pay announcement

Trainor said authentication stays foundational, and that the mistake is treating the model as the system of record or as the processing engine. He called the working future governed agency. Magennis said he would start with bounded uses, clear spending limits and clear accountability. Mike Storiale, senior vice president for AI technology and transformation at Synchrony, said intent and authentication both have to be there, and that many building blocks for trust have to land first.

Waller, speaking at 3:00 p.m. EDT on September 29, 2026, at Sibos in Miami, split the market in two. In the assisted model, the agent searches and the buyer still pays. In the delegated model, the buyer grants authority, sets guardrails, and the agent shops and pays. Consumer commerce is the first wave, with assistance before full delegation. Business buying may fit agents well because of approved suppliers and budgets, but the tickets are larger, so a bad action costs more, and firms still pay across ACH, wire, instant rails and cards.

The question shifts from proving that a buyer is an authorized payer to proving that an agent has the authority to pay on the buyer’s behalf.

Christopher J. Waller, Member, Board of Governors of the Federal Reserve System, in Payments in the Age of AI Agents, Sibos 2026, Miami

His other two walls are liability (“Who is on the hook if an agent makes the wrong purchase?”) and fraud models still tuned to human patterns. He asked whether the market will settle on platform-specific rules or on rules that work across stores, agent apps and payment methods. The full text of Payments in the Age of AI Agents is the clearest official map of those walls. Todd Kennedy, executive vice president for payments strategy at Capital One, said banks need more visibility into how a purchase is started, authorized and finished. Suzanne Sando, lead analyst of fraud management at Javelin Strategy & Research, said fraud systems must learn whether an agent is trusted, authorized and inside the buyer’s intent.

Unfindable SKUs Stop the Agent Cold

Storiale’s warning is about search that no longer looks like search. “If you’re not focused on that today, if you’re not prepping your brand to show up in these platforms, if you’re showing up 10th there, but first in traditional search, that’s the place you got to be putting your priorities,” he said. Hiatt was shorter: “If the SKUs aren’t available and findable, nothing else matters.”

Waller made the same point from the central bank side. Agents need live knowledge of stock. Big retailers are feeding catalogs to third-party agents or running their own shopping agents. Smaller merchants are being pulled along by commerce platforms that standardize product data so leading agent interfaces can see it. A brand that wins Google and loses the agent feed is invisible at the moment the cart is built.

Ng said the quiet substrate is structured invoices, payment history, counterparty files, policies and rail data. Combined, those records can help an agent choose how to pay, then help finance teams match and stay inside the rules. That invoice path is also where enterprise agents that copy invoice-fraud loops already do damage when they repeat a bad pattern at machine speed. Dave Ruda, vice president of product at Billtrust, said his first pick is the global compliance piece. “If you can’t get that done, then you’re not going to make it to the second round.”

Talk to the Agent, Skip the Screen

Ruda described the emerging model as headless financial software. “You don’t need front end to do anything. You just need to be able to talk to the agent,” he said. “Think about all the tasks you do now being spun up by agents and being automatically executed. That’s pretty powerful and very real.”

A household could let an agent reorder goods under a cap, or book travel inside set rules, without tapping approve on every charge. Magennis still wants those uses bounded. The instruction sounds simple. The stack under it has to keep identity, intent, merchant, credential, spending limit and the payment itself in one chain. Break any link and the issuer sees a strange charge with no story attached.

Trainor’s reach problem sits here too. Billing in one system, credentials in another, and a chat agent in a third is how you get a fluent answer and a failed payment. Model Context Protocol and similar hooks let software talk to tools without a person clicking through screens. For bill-pay and collections shops, that is the point of the unglamorous rebuild: agents can hit the ledger, not a web form.

Whoever Holds Intent Owns the Transaction

A bank, network, issuer, buy-now-pay-later firm or merchant that waits for the buy button may find the consequential choice already made in an agent somewhere else. Hiatt’s routing layer, Ng’s invoice substrate, Storiale’s catalog rank, Trainor’s reach, Ruda’s compliance gate and Magennis’s machine-readable chain are the same bet from six desks: the swipe is the end of the story.

Mastercard’s U.S. test of the agent score is the tell that issuers still cannot see the software in the middle. Until they can, most carts stay human, most roadmaps stay full, and the firms that own permission, catalogs and routing collect the value AI was supposed to skip.