An AI Agent Just Completed a Real Payment on Visa's Rails. A 140-Company Consortium Thinks the Rails Are the Problem.

On July 6, an AI agent spent money on a human's behalf using a real Visa card, through a real Spanish bank, on infrastructure nobody had to invent. That's the detail worth sitting with: the fight over who controls agentic payments was never going to be won by whoever built the smartest model. It was always going to be won by whoever already owns the rails the money moves on.

CaixaBank and Visa completed the transaction as a live pilot, not a demo — an AI agent initiating a purchase on a cardholder's behalf, processed through Visa's existing tokenization, identity verification, and fraud-detection systems rather than a newly built protocol. CaixaBank, which serves roughly 12 million cardholders, is running the pilot inside a €5 billion technology strategy stretching through 2027, built around deploying generative AI at bank scale. The point of the exercise was to prove that machine-initiated commerce doesn't require the financial system to be rebuilt. It requires the financial system to extend trust to a new kind of buyer.

Days earlier, Visa had quietly hedged that exact bet. It joined more than 140 companies — Mastercard, Stripe, BlackRock, Google, Coinbase, IBM, BNY, Standard Chartered, and others — in launching Open USD, a stablecoin engineered so no single company issues it and reserve yield gets shared across partner firms instead of captured by one issuer. Circle's stock fell as much as 17% on the announcement before rebounding roughly 7% on July 6, as markets worked out how seriously to take a coalition that includes the two largest card networks on earth.

Read side by side, these aren't contradictory moves. They're the same company betting on two different futures at once. The CaixaBank pilot preserves Visa's position as the toll booth every card transaction already passes through — just with a consent layer bolted on for non-human buyers. Open USD is the opposite wager: that payment rails eventually migrate to blockchain-native settlement anyway, and it's better to co-own the winning structure than get disintermediated by whichever stablecoin issuer gets there first.

Stripe is running a third strategy in the same window — a Machine Payments Protocol and Agentic Commerce Suite that opens its Link wallet to AI agents and is wiring agentic checkout directly into Google Gemini and Facebook ad surfaces, on top of token-based billing infrastructure for pay-as-you-consume pricing. Where Visa is defending a network and OUSD is rebuilding stablecoin economics, Stripe is betting the winning position is the API layer that sits above both and talks to whichever rail is cheapest that day.

None of these three bets actually answers the question that matters most to anyone underwriting this category. The interesting problem was never whether an AI agent could technically complete a transaction — CaixaBank just showed that it can, on infrastructure banks have run for decades. The unresolved problem is consent and liability: who verifies an agent was actually authorized to spend, who is on the hook when it buys the wrong thing or gets manipulated into buying anything at all, and who eats the fraud loss when the buyer isn't a person who can be asked to explain themselves. That's a thinner, less glamorous layer than a stablecoin consortium or a card network's press release, and right now it's underpriced relative to every application-layer AI wrapper riding on someone else's payment rails.

CaixaBank's pilot worked cleanly because a human still explicitly authorized the purchase before the agent acted. The infrastructure fight only gets interesting once agents start making purchase decisions the bank never explicitly underwrote for — and none of the three approaches racing to own this moment, card rails, shared-governance stablecoins, or API middleware, have actually said who's liable when that happens.