On June 10, 2026, Visa and OpenAI announced that tokenized Visa credentials can power agent-initiated checkout inside ChatGPT and Codex, and Mastercard launched Agent Pay for Machines the same day. The card networks are giving AI agents a sanctioned, scoped way to pay, which changes how an AI buys from your store.
- An agent token is a network token scoped to one agent, one merchant scope, and a consent policy with spending limits, not your raw card number.
- The goal is to make agent purchases safer and clearer than an agent typing a card number into a form, with verified agent identity and real-time checks.
- How chargebacks, liability, and authorization rates settle for agent-initiated, tokenized orders is still being worked out.
- The merchant move is to stay on native payments, expose your catalog to agents, and track agent orders separately once you can see them.
On June 10, 2026, the card networks made the same bet on the same day. Visa and OpenAI announced that tokenized Visa credentials can power agent-initiated checkout inside ChatGPT and OpenAI's Codex coding agent. Hours later, at a separate event, Mastercard launched Agent Pay for Machines, its own framework for letting verified AI agents pay. Two of the largest payment networks in the world looked at agentic commerce and decided the missing piece was a sanctioned way for an agent to hold a credential and check out. Now they are racing to supply it.
I have run the payments side of consumer brands, where authorization rate, fraud rate, and chargeback ratio were not abstract. They were the difference between a healthy month and a painful one. So when the networks move on something this foundational, I read it as an operator, not a futurist. The question I care about is narrow and practical: when an AI agent buys from my store using one of these tokens, what changes about whether the payment goes through, who eats the fraud, and what the order actually costs me. That is what this post is about.
This is the payments layer of the same shift I have been tracking across the blog. The discovery side, the agent that finds and recommends you, I covered in the agentic commerce primer for Shopify brands. This is the other half: once an agent has chosen your product, how it pays. Tokenized agent credentials are the rails that make agent checkout real instead of a demo, and the networks just started laying them.
Two networks moved
on the same day,
on the same idea.
Start with the facts, because the framing matters. Visa and OpenAI announced their collaboration on June 10, 2026: Visa's tokenization, agent identification, and fraud-monitoring infrastructure routed into OpenAI's consumer and developer systems, so a tokenized Visa credential can power an agent-initiated checkout inside ChatGPT and Codex. Transactions run inside user-defined controls, spending limits, merchant categories, and approval thresholds, enforced before authorization. As of mid-June it began partner-limited, with no published consumer go-live date.
Mastercard launched Agent Pay for Machines the same day. It lets verified agents transact using what it calls Agentic Tokens, which bind a tokenized card credential to a specific agent, a specific merchant scope, and a specific consent policy. More than thirty-five partners were named, including Stripe, Adyen, Checkout.com, and a set of crypto-native names, and the service is built to settle across cards, bank accounts, and stablecoins. Both networks, on the same day, shipped the same core idea: a credential an agent can hold that is not your raw card.
"When Visa and Mastercard move on the same idea on the same day, that is not a trend to watch. That is the rails being poured."
When the two networks that sit under most of the world's card volume move in lockstep like this, it tells you the direction is set even if the timeline is not. Agent-initiated payments are going to have a sanctioned path. The open questions are when it reaches your checkout and on what terms, not whether it arrives. That is a very different posture than the discovery-layer fights, where the winner is still genuinely uncertain.
A tokenized agent
credential, in plain
operator language.
Here is the whole idea without the jargon. Today, if an agent wanted to buy from you, the cleanest path was to type your customer's actual card number into your checkout, which is exactly the scenario that makes a fraud team sweat: a bot holding a raw card. A tokenized agent credential replaces that. The network issues a token that stands in for the card, and it scopes that token to one agent, one merchant context, and a consent policy with limits and approval rules baked in.
That scoping is the point. If the agent token leaks or the agent misbehaves, the exposure is bounded to that token's narrow scope instead of your customer's entire card. The network can verify which agent is transacting, check it against the consent policy in real time, and run its fraud monitoring on a transaction it actually understands, rather than a mystery checkout from an unidentified browser. It is the same network tokenization that already sits under Apple Pay and Google Pay, extended with agent identity and a consent policy.
An agent token is a stand-in for the card that only works for one agent, at a defined set of merchants, within rules the customer set. The network knows which agent is paying and can check the rules before it approves. That is meant to be safer than an AI typing a real card number into a form, and clearer for everyone in the chain about who authorized the purchase.
If you want the layer that sits between this and your store, the protocols that govern how an agent reaches your checkout in the first place are in the UCP versus ACP comparison, and how you expose products for an agent to buy is in the catalog API and AI agents breakdown. The token is the payment instrument. Those are the roads it travels on.
Visa with OpenAI,
Mastercard with its
own machine rails.
The two announcements rhyme, but they are not identical, and the differences hint at where each network is placing its bet. Visa went deep with one large AI partner, embedding directly into OpenAI's surfaces. Mastercard went broad, shipping a framework with dozens of partners and multi-rail settlement, including stablecoins, aimed explicitly at machine-to-machine and high-velocity transactions. One is a partnership, the other is a platform play.
| Dimension | Visa + OpenAI | Mastercard Agent Pay for Machines |
|---|---|---|
Shape of the move | Deep integration with one AI partner | Broad framework, 35-plus partners |
Where it lives | Tokenized Visa creds in ChatGPT and Codex | Agentic Tokens across many agents and rails |
Token approach | Network token scoped to the agent | Agentic Token bound to agent, merchant, policy |
Settlement | Card rails | Multi-rail: cards, bank, stablecoins |
Status at launch | Partner-limited, no consumer date | Launched with partner roster |
For a DTC brand, you do not have to pick a side, and you should not try. Both run on network tokenization you already inherit through your payment platform. The useful read of the comparison is what it signals: the entire card industry has decided agent payments need a sanctioned credential, and it is being built two ways at once. Your job is not to bet on Visa or Mastercard. It is to be on payment infrastructure that will support whichever path reaches your checkout, which for most Shopify merchants means staying native.
Does this make fraud
better or worse for
your store?
This is the question I get first from operators, and the honest answer is: probably better in theory, still unsettled in practice. The theory is sound. A tokenized agent credential carries verified agent identity, a consent policy, and real-time authorization, which is a far better fraud posture than an unidentified agent jamming a raw card into your form. The network can see who is paying and check the rules. That should reduce a whole class of bot fraud, not create it.
What is genuinely unsettled is liability. When an agent initiates a tokenized purchase and the customer later disputes it, who eats the chargeback, and under what rules? Is an agent-initiated transaction treated as card-present-like, with the verified token shifting liability, or does it land on the merchant the way card-not-present often does? The networks are clearly building toward a cleaner answer, but the dispute and liability mechanics for this specific transaction type are still being worked out, and that is the line a merchant should watch most closely.
My advice is to treat agent-token orders as a new cohort the moment you can identify them, and watch their authorization, fraud, and dispute rates against the rest of your book before you assume anything. Do not let a vendor tell you it is automatically safer or automatically riskier. Measure it. This is the same discipline I push on every margin and risk line, the kind of teardown I walk through in the DTC profitability teardown.
What agent payments
could do to your
unit economics.
Payments are a margin line, so any change to how a transaction is initiated is a P&L question, not just a tech one. There are two ways agent payments touch your economics, and they pull in opposite directions. On the upside, a cleaner, verified, tokenized transaction can mean better authorization rates and fewer false declines, and every recovered legitimate sale is pure margin you were otherwise losing at the door.
On the other side sits the toll-booth risk. If the agent layer becomes a powerful intermediary, the companies that own the agent and the payment rails will look to take a cut, the same way every layer that sits between you and your customer eventually does. Whether that cut is real, and how big, is not knowable yet, but it is the thing to model before you build a business that depends on agent-driven volume. I think through that exact dynamic in the contribution margin breakdown, because a new intermediary fee is just another line above contribution.
Run agent orders through the same contribution-margin math as any channel. What is the effective cost of the sale, including any agent or token fee, the fraud and dispute load, and the authorization lift? If agent orders convert better and authorize cleaner, they may be your best channel. If a toll appears, you want to have priced for it. Either way, you only know by measuring the cohort.
The point is not to fear agent payments or to cheer them. It is to refuse to treat them as free or as automatically better. They are a new way money moves into your store, with their own authorization behavior, fraud profile, and potential fee structure. Put them in the model as their own line, and let the data tell you whether they help or hurt your contribution margin instead of guessing.
Three moves while
the agent-payments
rails are still wet.
This capability is partner-limited and early, so the wrong move is to re-architect your checkout for it today. The right move is to position so you inherit it cleanly and can read it honestly when it arrives. Three things do that.
Visa, OpenAI, and Mastercard moving on the same day tells you the agent-payments rails are being poured. It does not tell you the road is open yet. For a DTC brand, that is the right time to position and the wrong time to over-build. Stay on payment infrastructure that will carry the new credentials, make your store buyable by an agent, and get ready to measure agent orders as their own cohort the moment they show up. The networks are deciding how an AI pays you. Your job is to be ready to read whether it pays you well.
What brands ask me
about agent tokens
and checkout.
On June 10, 2026, Visa and OpenAI announced that tokenized Visa credentials can power agent-initiated checkout inside ChatGPT and OpenAI's Codex. The deal routes Visa's tokenization, agent identification, and fraud-monitoring into OpenAI's systems. It is an announced capability that began partner-limited, with no public consumer go-live date when it was unveiled. The discovery side of this shift is in the agentic commerce primer.
It is a payment token scoped to a specific agent rather than to your raw card. A network token replaces the card number and is bound to one agent, a merchant scope, and a consent policy with spending limits and approval rules. If that token is compromised, exposure is limited to that one agent's scope instead of your whole card. The roads it travels on are in the catalog API breakdown.
Mastercard launched Agent Pay for Machines on June 10, 2026, the same day as the Visa-OpenAI news. It lets verified AI agents transact using Agentic Tokens that bind a card credential to a specific agent, merchant scope, and consent policy, built on the same MDES tokenization behind Apple Pay and Google Pay, and settling across cards, bank accounts, and stablecoins. The protocol layer beneath it is in the UCP versus ACP comparison.
The goal is to make agent purchases safer and clearer than an agent typing a raw card number, because the token carries verified agent identity, real-time authorization, and a consent policy the network can check. How chargebacks and liability land for an agent-initiated, tokenized transaction is still settling. Watch your processor's terms and your authorization and dispute rates as these credentials roll out. Read them like any margin line, the way I do in the profitability teardown.
Do not re-architect checkout for a partner-limited capability. Stay on your platform's native payments so you inherit agent-token support as it ships, expose your catalog so agents can find and buy you, and track agent-initiated orders separately once you can see them so you can read their authorization, fraud, and margin behavior against the rest of your book. The contribution-margin frame for it is in the contribution margin post.
Reading what agent payments do to your checkout and margin?
I ran the payments, fraud, and authorization side of consumer brands at nine-figure scale, where a half-point of authorization rate moved the month. I help DTC teams and Shopify merchants read the agent-payments shift against their actual P&L and decide what to change and what to leave alone. If this is on your roadmap, that is the work I do.
Start a conversation See the case studies →A note on sources: The June 10, 2026 Visa and OpenAI collaboration, with tokenized Visa credentials powering agent-initiated checkout in ChatGPT and Codex and the launch being partner-limited, is from PYMNTS, Visa, and Digital Commerce 360. Mastercard's Agent Pay for Machines, its Agentic Tokens, MDES basis, multi-rail settlement, and partner roster are from Mastercard and Fortune. The read on fraud liability, unit economics, and the merchant playbook is mine, from running the payments side of consumer brands at scale.