Visa, Mastercard and Ant's AI-payments pact is a fight over who keeps the toll booth

Three of the biggest names in payments agreed this week that software should be allowed to shop on their networks — and agreed on a shared rulebook for deciding which software to trust. The obvious way to read it is as a growth story: AI agents handling $3 trillion to $5 trillion of consumer commerce by 2030. The more revealing way is to ask what the rulebook actually preserves.

The pact, announced September 10 by Visa, Mastercard and Ant International, is called a "Know-Your-Agent" interoperability framework, and it was convened through Singapore's BuildFin.ai platform. Here is the setup it is trying to solve. An AI agent — software that books a flight, reorders groceries or negotiates a bill on your behalf — is not a customer in any sense a payments company has known. It has no card to tap. It has no identity to verify. And, most importantly, it cannot be held responsible for anything it does.

When you tap a card today, you are the pool of liability. You carry the fraud, the disputes, the returns. That human-in-the-loop is what lets Visa and Mastercard clear a transaction and trust that the bill will be paid. An agent that acts for you removes that person from the moment of purchase — and if the agent acts badly, there is no one left to hold accountable. That is a direct threat to the economics of the card networks, because it opens the door to a world where machine purchases route around their rails entirely, settling through wallets, closed platforms or stablecoins where no one's card needs to be present.

The framework's answer is its defining feature: "cross-network operator traceability." Every agent action is linked back to a validated cardholder, business or organization — a real human or legal entity that absorbs the consequences. In plain terms, the machine gets to spend, but a person stays on the hook. That accountability is the one function in a purchase that software cannot perform for itself, and it happens to be exactly the gatekeeping role the networks already own.

This is worth sitting with, because it reframes what the announcement is actually selling. It is not primarily a bet that AI agents will be huge. It is a defensive bet that the trust layer — identity, verification, fraud, accountability — must stay human, and that it should stay in the hands of the networks. Their own financials show why they want that. Mastercard's fastest-growing line is precisely this: value-added services for security, authentication and digital identity, which grew 20% year over year to $3.8 billion in its latest quarter, faster than its core switched transactions. The AI-agent framework is, in effect, an argument that this franchise becomes more valuable, not less, as machines do more of the buying.

A partnership of rivals

The striking thing about the deal is that it is a collaboration among three competitors. Each company already shipped its own agent protocol over the past year — Visa's Trusted Agent Protocol, Mastercard's Verifiable Intent, and Ant's Agentic Mobile Protocol. A shared rulebook lets an agent registered on one network move across the others without re-registering, which grows the category and lowers integration costs for everyone. That is the cooperative logic.

But cooperation sets only a floor. Each partner still hopes its own protocol becomes the layer where value actually accrues, and each keeps its own risk controls and its own toll at checkout. This is standard platform behavior: share enough to grow the pie, compete to own the point where economics concentrate. An "open, interoperable" standard is as much a shield against being seen as a gatekeeper as it is a genuine service — and it lets the three giants set the rules of machine commerce before anyone else does.

The wild card at the table is Ant International, and it deserves more attention than a joint-press-release mention. Ant is the only partner that does not run a card rail. It runs digital wallets — its Alipay+ gateway connects merchants to more than 50 wallets and banking apps reaching some two billion user accounts — and it has been steadily routing transactions around cards for most of its existence. It was the disruptor of card payments in China; giving it a seat to help write the rulebook for agent payments is both insurance for it and a reminder, for Visa and Mastercard shareholders, that the threat to the networks is not hypothetical. It is sitting across the table.

What you can't yet conclude

Set the forecast aside before judging this. The $3–5 trillion figure is a projection the companies are citing from McKinsey, not a booking, and it describes a 2030 world on an optimistic scenario. A standards agreement signs no customers, moves no money and produces no revenue line. For Visa and Mastercard, both of which just grew net revenue 14% in their latest quarters on payments volume that already exists, this pact changes none of the current numbers.

What it does is buy a position. Think of it as low-cost option: Visa and Mastercard are spending engineering and a shared press release to ensure that when machine commerce actually scales, the transactions touch a rail that collects a take, and the trust function that underwrites them stays in their hands rather than drifting to a platform, a wallet or a settlement layer they don't control. Their bet is that the more agents transact, the more valuable the one thing they own — accountability for a buyer that can't be blamed — becomes.

The sharper risk is the one buried in the framework's own design. The entire point is open interoperability and shared certification — which is another way of saying the trust layer could be commoditized. If every network vouches for the same agents under the same rules, then no single gatekeeper can extract durable economics from the job it is doing; the take could be competed down to nothing. In that world the winners are whichever rails the agent transactions actually choose to settle on, whether that turns out to be a card network, Ant's wallets, or a cheaper settlement layer underneath both. That is the question worth watching — not the size of the forecast, but where the first real machine-originated dollars end up settling, and whether the network that vouches for them is the one that gets paid.