If agentic commerce is AI agents buying on a shopper's behalf, someone still has to move the money. Securely, and without handing a bot your card number. As of mid-2026, the card networks are answering that with tokenised, scope-limited credentials: a payment token bound to a specific agent, purpose, and spending limit, so an agent can pay without ever holding your real card details.
Who actually moves the money in agentic checkout?
The payment networks do. The same Visa and Mastercard rails behind ordinary card payments, extended so an AI agent can be the one initiating the charge. According to reporting in June 2026 (AP and others), Visa embedded its payment network into ChatGPT, giving the assistant access to Visa's tokenisation and fraud infrastructure for agent-initiated purchases. The agent researches, compares, and triggers payment; the network clears it.
This is a distinct layer from the checkout protocols. Standards like the Agentic Commerce Protocol (ACP) define the shared language for an agent and merchant to transact; the card networks provide the money movement underneath. A complete agentic purchase needs both.
How do tokenised agent credentials work?
Tokenisation replaces your sensitive card data with a network token, but for agents, the token is also scoped. Per June 2026 reporting on Visa's approach, a token is bound to a specific agent and use case, with limits enforced at the network:
- A token issued for a grocery shopping agent cannot be used to book travel.
- A token carrying a $200 limit cannot authorise a $500 transaction.
- The agent transacts with the token; your real card number is never exposed to it.
The effect is least-privilege payments: the agent gets exactly enough authority to do the one job, and no more. That containment is what makes delegating a purchase to software tolerable in the first place.
What did Visa and OpenAI announce?
Reporting in June 2026 described Visa embedding its payment network directly into ChatGPT, framed by Visa as part of "agentic commerce". AI systems that don't just suggest products but research items, compare prices, and help complete payment. Notably, the same reporting indicated that at launch, Visa requires human approval for most transactions: ChatGPT sends the shopper a notification to confirm each purchase before it completes. Treat the specifics as point-in-time; the launch posture and limits are early and will evolve.
Tokenised agent credentials make the money safe to move. They don't make your product the one the agent decides to buy, that's still earned upstream.
How is Mastercard's approach different?
Mastercard is building a parallel capability rather than ceding the lane. Its framework, reported as Mastercard Agent Pay, lets verified AI agents transact on a consumer's behalf using Agentic Tokens: described as an extension of the Mastercard Digital Enablement Service (MDES), the same tokenisation backbone behind its existing digital payments. Reporting in June 2026 placed more of Mastercard's early emphasis on business procurement use cases. The shared idea across both networks: a verified agent, a scoped token, and controls at the network layer.
| Network | Mechanism (as reported, mid-2026) | Early emphasis |
|---|---|---|
| Visa | Payment network embedded in ChatGPT; tokenised, scope-limited credentials; human approval for most transactions at launch | Consumer shopping inside the assistant |
| Mastercard | Agent Pay framework; Agentic Tokens extending MDES | Verified-agent transactions, incl. business procurement |
Names, scopes, and availability are moving quickly. The durable takeaway is the pattern, not any one product detail.
Where payments sit vs. being chosen
Here's the distinction brands keep missing in the launch noise: the payment rail is distribution and settlement, not selection. It moves the money once an agent has already decided to buy your product. It does nothing to make your product the one the agent chooses.
That choice still turns on the layer above the rails. Catalog richness, entity strength, reviews, and how consistently other sources corroborate your claims. It's the same answer-engine optimisation that decides whether you're surfaced at all. Wiring up a flawless agentic payment path for a product no agent recommends is a fast checkout to nowhere.
So treat the payments news as confirmation the transaction layer is real, and then spend your effort where selection is won. Knowing whether agents and answer engines actually surface and pick your products, as these rails mature, is exactly what Buffy Intel measures. Questions: [email protected].