Agentic Commerce Infrastructure: The Rails Decide Who Wins

Somewhere in the last two weeks, agentic commerce quietly graduated from conference slide to shipping product. Worldline, ING and Visa ran a live agentic payment in Europe. Cross River and Stripe lined up card issuance built specifically for agents. Microsoft wired a Model Context Protocol server straight into Dynamics 365 Commerce. And Juniper Research dropped the number everyone will be quoting for the next year: 1.3 billion agentic commerce users by 2031, up from under 300 million today.

That is a lot of agents holding a lot of wallets. Juniper also put transaction value on a path from 8 billion dollars in 2026 to 3.5 trillion by 2031, which is either the future of buying or the single most optimistic slide of the decade, depending on exactly one thing. Not the AI. The agentic commerce infrastructure underneath it.

Key Takeaways

The AI was never the hard part

The most honest line in the whole news cycle came from Adyen. Speaking to PYMNTS, the company's agentic commerce lead rated the market at roughly 0.5 on a five-point maturity scale, expected to triple inside a year and still nowhere near the middle. The point underneath the number is the one that matters: the difficult work was never teaching a model to shop. It was the plumbing beneath the checkout. Machine-readable catalogs. Trust. Fraud. Liability. A pile of competing protocols that nobody has fully reconciled.

Read that back slowly, because it reframes the entire category. Everyone spent two years arguing about which model reasons best. Meanwhile the actual bottleneck sat one layer down, in the boring, unglamorous, absolutely load-bearing question of how a piece of software proves who it is and moves money without setting off every alarm in the building.

An agent that can reason brilliantly and cannot pay safely is a very expensive window shopper.

Read the pattern hiding in the headlines

Line the announcements up next to each other and something obvious appears. They are not competing on intelligence. They are all quietly building the same four things.

When Worldline, ING and Visa completed their live agentic payment, the interesting detail was not that an agent paid. It was how. A person set the conditions of the purchase in advance, the agent acted within those parameters, the issuer kept authorization control, and the whole thing left a verifiable record of what was authorized, by whom, and under what conditions.

When Cross River and Stripe described their issuance work, they framed the core challenge as trust: proving that a transaction reflects genuine intent, carried out by a verified agent, inside the scope the user actually authorized. Single-use credentials, tied to a user's permission, bounded by merchant and amount.

When Microsoft shipped its MCP server for Dynamics 365 Commerce, the framing was that an agent is only as useful as the systems it can reach, so a platform exposes its capabilities once and agents call them in real time, with guardrails set before anything goes live.

Different companies. Different stacks. Same shopping list every single time:

• A way to know which agent is acting, and on whose behalf.
• A wallet or credential the agent can actually use.
• Payment rails that clear the transaction.
• Policy that bounds what the agent is allowed to do, with a record to prove it did.

Nobody in this cycle is fighting over the model. They are all fighting over the rails.

The four things an agent needs to transact

This is the part where a category stops being a trend and starts being infrastructure. If every serious player is converging on the same four requirements, those four requirements are the product. At Abstraxn we build to exactly that map.

Identity, wallet, payments, policy. The market spent its attention on the model. The value accrued to the four primitives underneath.

The card caveat everyone should read twice

One more thing from Juniper, and it is the most useful sentence in the report for anyone building here. Cards took an early lead in agentic commerce, with the networks running most of the first pilots. Then the researchers added the warning: card domination is not in the market's best interest, and platforms that fail to support local payment methods, digital wallets and account-to-account transfers will cap their own growth.

Translation for builders: do not hard-wire your agent economy to one payment method. The purchases are going to happen across cards, wallets, stablecoins and rails that have not standardized yet, on more than one chain. Infrastructure that assumes a single method is infrastructure with an expiry date. The rails that win are the ones that stay method agnostic and chain agnostic from day one, which is precisely the design constraint we started from rather than the one we are now retrofitting.

The window is open, briefly

Adyen's other point deserves a line of its own: this is the experimentation window, and waiting for the category to settle is its own kind of decision. The agents are getting good. The forecasts are absurd in the way that early internet forecasts were absurd, right up until they were conservative. The teams that wire up identity, wallets, payments and policy now are the ones who will be collecting fees while everyone else is still reading the standards documents.

You do not have to rebuild any of this. That is the entire point of infrastructure.

FAQ