Two Networks Opened Their Rails to Agents in Two Days. Neither Said Who Pays.
Visa and Mastercard opened their rails to agents twenty-four hours apart. Neither touched the rulebook.
Visa and Mastercard both shipped credentials and controls this week. The rulebook — the part only they can write — went untouched.
On Tuesday, Mastercard launched Agent Pay. On Wednesday, Visa launched Intelligent Commerce. Twenty-four hours apart, the two companies that between them define how most of the world's card transactions are allowed to work both announced that AI agents may now transact on their networks.
I read both releases the same evening, looking for one specific thing. It is not in either of them.
Neither network said what happens when an agent buys something the cardholder did not want.
That is not a gap in a press release. It is a gap in the rulebook — and the rulebook is the one thing in this entire story that nobody but Visa and Mastercard can write.
What each one shipped
Mastercard, 29 April. Agent Pay, introduced as an agentic payments program, built on Mastercard Agentic Tokens. The tokens extend the same technology that sits under contactless mobile payments and under programmable arrangements like subscriptions and recurring bills. Launch partners are Microsoft for consumer use cases, IBM for business-to-business, and Braintree and Checkout.com on the tokenization side. Jorn Lambert, the chief product officer, framed it as initial steps in redefining commerce in the AI era, and pointed at new merchant interfaces meant to distinguish trusted agents from bad actors. The company's language for the destination is an agentic commerce future where consumers and businesses transact with trust, security and control.
Visa, 30 April. Intelligent Commerce, announced at Visa's product drop, opening the network to developers building agents. The centerpiece is what Visa calls AI-Ready Cards: tokenized credentials that replace card details and confirm that a consumer's chosen agent is authorized to act on their behalf. Consumers set spending limits and conditions. Commerce signals flow to Visa in real time, which the release says lets Visa apply transaction controls and help to manage disputes. The partner list is long and notable: Anthropic, IBM, Microsoft, Mistral, OpenAI, Perplexity, Samsung and Stripe.
One sentence in Visa's release is worth copying out exactly, because it is the clearest statement of the control model either company made: "Only the consumer can instruct the agent on what to do and when to activate a payment credential."
Hold on to that sentence. We are coming back to it.
They shipped the same thing, and it is the easy half
Strip the branding and the two announcements are architecturally almost identical.
Both replace the card number with a token specific to the agent. Both let the cardholder set limits in advance — amounts, categories, conditions. Both keep a record of the consumer's authorization of that particular agent. Both promise merchants a way to tell a sanctioned agent from an unsanctioned one.
This is good work and I do not want to be glib about it. Tokenization is the single most useful thing that has happened to card security in fifteen years, and extending it to a new class of actor is exactly the right instinct. The networks are doing what networks are genuinely excellent at: issuing credentials and attaching conditions to them.
But notice the shape of the control. It is all pre-authorization. The cardholder decides, in advance, what the agent is permitted to do. Then the agent goes away and does it — minutes later, or days later, at a merchant the cardholder has never dealt with, with no human present at the moment of purchase.
We have a name for a transaction where a credential is stored and used later under a prior arrangement. It is not a new category. It has rules, reason codes and liability allocations, and every one of them was written on an assumption that is about to stop being true: that the party initiating the transaction is the merchant, and that the merchant already has a relationship with the cardholder.
An agent is neither of those things. It is a third party, acting across merchants, under a mandate the cardholder gave to somebody else entirely.
The half nobody shipped
Here is the thing a network owns that nothing else in payments can replicate.
Not tokenization. Tokenization is excellent engineering, but it is engineering — Stripe does it, Apple does it, every large merchant does it on stored cards. Not the credential, not the API, not the developer program.
What a network owns, uniquely and absolutely, is the rulebook: the document that says which category a transaction falls into, what evidence each party may bring, and who absorbs the loss when the two sides disagree. That is not a product. It is closer to private law, and Visa and Mastercard are the only bodies that can amend it.
Both companies this week shipped the part that is merely difficult. Neither touched the part that is theirs alone.
Look again at Visa's sentence. Only the consumer can instruct the agent on what to do and when to activate a payment credential. It is a careful, accurate statement about authorization. It says nothing at all about what happens when the instruction was given and the outcome was still wrong.
Because an agent purchase can fail in at least three distinct ways, and today they all land in the same bucket.
The agent was not authorized. Somebody compromised the credential and ran an agent the cardholder never approved. This one is easy, and the existing rules handle it: it is fraud, and it resolves the way card-absent fraud resolves.
The agent was authorized and bought the wrong thing. The cardholder told it to buy a birthday gift under two hundred dollars. It bought a birthday gift under two hundred dollars. The cardholder hates it, or it was the wrong size, or it was for the wrong person. The instruction was followed. The outcome is unacceptable. There is no reason code for the machine did what I said and I did not mean it, and no framework for deciding whether that belongs to the agent platform, the merchant, or the cardholder.
The agent was authorized, bought exactly the right thing, and the human changed their mind. This is ordinary first-party misuse, and we have dealt with it for decades. But agents hand it something it never had before: a respectable alibi. I didn't buy that, my agent did is an explanation that sounds technical, sounds plausible, and is nearly impossible to disprove from the issuer's side.
That third case is the one that should be keeping issuing risk teams awake, and it is the one nobody is discussing, because it is not a technology problem and has no product launch attached to it.
Two different bets, hiding in the partner lists
The partner lists are the most revealing part of both announcements, and they do not match.
Visa signed the model companies: Anthropic, OpenAI, Mistral, Perplexity, alongside Microsoft, Samsung, IBM and Stripe. That is a bet that the agent platforms are the kingmakers — that whoever builds the agent chooses the rails, and the way to win is to be the network the agent was built against.
Mastercard signed the acceptance side: Braintree and Checkout.com on tokenization, with Microsoft and IBM for use cases. That is a different bet — that agentic volume arrives through the existing processing stack, and the way to win is to make sure every gateway already speaks your token.
Both could turn out to be right, and they imply very different distribution. One pushes the new credential down from the agent; the other pushes it up from the acquirer. For those of us further down the chain the practical consequence is unglamorous but real: agent traffic will reach our BINs by more than one path, at different times, with different amounts of advance warning, depending on which bet matures first.
From a Mexico City desk
Neither program launches here this quarter, and that is the least interesting fact about them.
Our cardholders will meet agents from the other side of the border, the same way they meet everything else — as a cross-border card-absent transaction that arrives on our BIN with no label explaining what it is. We will receive the category before we can offer it. That is simply the order in which things reach a market like ours.
But there is a second-order problem here that I think is specific to our region and worth naming early.
The entire control model announced this week rests on the cardholder configuring something. Spending limits. Categories. Conditions. The consumer sets the guardrails and the agent operates inside them.
Now ask, honestly, what proportion of a portfolio in this region has ever opened the app and configured a card control. Self-service controls are available, they are promoted, and they are used by a small minority. The majority of cardholders will never set an agent limit — which means the limits governing agent spending on our cards will be whatever default somebody else chose: the agent platform, the wallet, the party that issued the token.
A control that nobody configures is not a control. It is a default with good public relations.
And then the part that arrives first, before any of the product work: the phone call. Somebody is going to call a contact centre in Guadalajara and say, in Spanish, that they did not make a purchase but their computer did. That call will arrive before the training material, before the reason code, before the policy. The script will be improvised by whoever answers.
That is worth getting ahead of, and it costs almost nothing. It is a page of guidance, not a project.
My judgment
This was a good week for the networks and a confusing one for everybody who has to operate underneath them.
They shipped the credential and kept the rulebook closed. That is the wrong order, and I do not think it is an accident. A credential is a product you can announce at a product drop. A rule change is a mandate with an effective date and compliance obligations, negotiated with thousands of members who will object to it. One of those is much easier to do in April.
But a credential without a rule is a transaction we can authorize and cannot adjudicate. We will approve these. We will not be able to classify them honestly. And when the first ones are disputed, we will resolve them by analogy to something they are not.
So my advice to anyone in an issuing or processing seat is to stop watching the product announcements. They will keep coming, they will be impressive, and they will not change your operations.
Watch the rules bulletins instead. The day a new initiation indicator appears in a mandate — the day there is a field that means an agent did this, under a mandate, with this much human involvement — that is the day agentic commerce becomes real for an issuer. Everything before it is a credential looking for a category.
That day has not come. Until it does, every agent purchase on your portfolio is being filed as something it is not.
Three questions for the people who own this
1. If an agent transaction arrived on your BIN tomorrow carrying one of these new tokens, would your authorization stack record that distinction anywhere a reporting query could find it? Or would it post as ordinary stored-credential e-commerce and vanish into the same bucket as every subscription you process?
2. What share of your cardholders has ever configured a spending control in your app? If the honest answer is a small minority, then whose defaults are about to govern agent spending on your portfolio — and did anyone at your institution agree to them?
3. When the first call comes in and the customer says their agent bought something they did not want, what does the person who answers actually say? Not the policy. The script. Does it exist?
Reply to this email with a line or two. The sharpest answers will shape the next issue.
Sources
Visa, 30 April 2025 — Visa Intelligent Commerce announced at Visa's product drop: AI-Ready Cards, tokenized credentials, consumer-set spending limits and conditions, real-time commerce signals, and the partner list. The quoted sentence on consumer instruction is from this release. Link
PYMNTS, 29 April 2025 — Mastercard debuts Agent Pay and Mastercard Agentic Tokens; launch partners; Jorn Lambert's comments on redefining commerce in the AI era and on distinguishing trusted agents from bad actors. Link
Neither announcement addresses liability allocation or dispute reason codes for agent-initiated transactions. The observation that this is missing is mine; the absence is verifiable in both sources.