Within a single day this week, the two largest card networks both declared that the coming economy of autonomous AI agents will settle on their rails. The competing announcements from Visa and Mastercard are not really about a new product. They are a claim to the most valuable position in agentic commerce: the trusted layer that decides whether a machine is allowed to spend money, and that carries the transaction when it does.
Visa used its Payments Forum in San Francisco to announce a partnership with OpenAI that lets AI agents make payments across OpenAI’s products, a deal it set out in its own announcement. The design keeps Visa in its familiar role. Agents transact with tokenized Visa credentials under real time authorization and fraud monitoring, and they operate inside user defined permissions such as spending caps, eligible merchant categories and required approvals. Marco Mahrus, OpenAI’s head of partnerships for commerce, framed agents as taking on a growing role in money movement and stressed keeping those transactions secure, transparent and under user control. The two companies said they would also explore enterprise uses, including OpenAI’s Codex coding agents buying inference, application programming interfaces or other developer services on their own within preset limits. That enterprise angle is easy to overlook and may matter most: a developer agent that can provision its own cloud and model resources, within a budget, is a procurement and treasury event, not a consumer checkout.
Two protocols, one land grab
Mastercard’s answer, launched the same day and described in trade coverage from Cryptopolitan, is more ambitious in scope and more explicit about machines. Agent Pay for Machines is an open protocol designed to let verified AI agents transact at machine speed, including micropayments worth fractions of a cent. It runs on what Mastercard calls Agentic Tokens, an extension of its existing Digital Enablement Service that binds a tokenized card credential to a specific agent, a specific merchant scope and a specific consent policy, a construct it brands Verifiable Intent. Crucially, settlement is multi rail, spanning cards, bank accounts and stablecoins, and agent permissions are recorded on public blockchains, initially Polygon, Solana and Base. The network launched with 31 partners, a list that tellingly mixes card world incumbents and crypto infrastructure: Adyen and Stripe alongside Coinbase, RippleX, the Solana Foundation, Polygon, Aave Labs, Anchorage Digital and Ant International.
The micropayment detail deserves emphasis because it points at use cases the card rails were never built for. Transactions worth fractions of a cent are uneconomic under traditional interchange, where fixed per transaction fees swamp the value being moved. An agent that pays per API call, per inference or per data lookup, thousands of times an hour, needs a settlement model with near zero marginal cost, which is exactly what stablecoin and on chain rails promise and what pushed Mastercard to make settlement multi rail rather than card only. Read that way, Agent Pay for Machines is less an extension of consumer payments than an attempt to claim the machine to machine economy, a flow of value between software that has no human in the loop at the moment of purchase and that traditional networks have no native way to price.
The strategic divergence is worth reading closely. Visa’s move is a defense of the existing model, extending tokenization and its trust framework to a new initiator, the agent. Mastercard’s is a hedge against that model, embracing stablecoins and on chain permissioning as settlement options rather than insisting everything routes through traditional card rails. Both reflect the same underlying anxiety. If agents become the entities that shop, negotiate and pay, the network that authenticates the agent and enforces its spending policy captures the durable value, and the one that merely moves the money becomes a commodity. Put bluntly, the networks are trying to avoid being disintermediated by the very automation their largest customers are racing to adopt, and the way to avoid it is to own the layer that says yes or no to a machine before any money moves.
The unsolved problems sit in the trust layer
None of this is settled infrastructure. Two incompatible agentic payment frameworks launching in one day is a standards fight waiting to happen, and merchants do not want to integrate twice. Liability is genuinely unresolved: when an agent buys the wrong thing, or is manipulated into spending through a prompt injection attack, the question of who eats the loss has no clean answer yet. Fraud monitoring built for human tempo has to be rebuilt for continuous high frequency automated commerce, where the behavioral signals that flag a suspicious human, hesitation, odd hours, unusual geography, mean nothing for software. And anchoring permissions to public blockchains introduces settlement and key management risks that compliance teams at regulated institutions will examine hard before they trust them with customer funds.
These moves also do not stand alone. They sit on top of a parallel push to make the money itself programmable, including Visa’s work building a technology layer that lets banks turn ordinary deposits into always on digital money. That matters because an agent settlement network is only as useful as the funds it can move, and tokenized deposits and stablecoins are what give an agent something to spend that clears instantly and around the clock. The card networks appear to be building both halves at once: the permission and trust layer that authorizes the agent, and the programmable money layer that the agent draws on. Whoever owns both ends of that pipe owns the economics of automated commerce, which is why neither Visa nor Mastercard can afford to cede the agent relationship to a model provider or a crypto network.
What it means for the finance leader
This is the moment the agent economy stops being a slide and becomes a procurement question. Treat it as infrastructure selection, not a pilot. The practical work is to map which of your payment flows could plausibly be initiated by an agent over the next two years, then evaluate the trust controls each network offers: how spending permissions are defined, how an agent is authenticated, how disputes and liability are handled, and whether stablecoin settlement is in scope. This is also a continuation of Visa moving to own the programmable-money layer, and the throughline is consistent: the networks intend to be the permission and trust layer for programmable money, whoever or whatever is doing the spending. Watch the standards question above all, because a two protocol world is unstable. Either the industry converges on a shared agentic payment standard, in which case early integration work transfers cleanly, or one network’s approach wins the merchant base and the other retreats, in which case betting heavily on the loser is costly. The pragmatic posture is to support the dominant flows you already run, keep agent payment pilots small and protocol agnostic, and insist any vendor can show how liability and dispute resolution work before a single autonomous transaction touches a customer account. The institutions that decide their agent payment strategy deliberately now will set the terms. The ones that wait will inherit whichever standard wins by default.
