Mastercard Will Let AI Agents Spend Your Money. It Also Wants to Decide, Later, Whether You Meant To.

Every few years a technology shows up that is supposed to cut the card networks out of the deal. Every few years the card networks end up in the middle of it anyway. AI agents are the newest version, and this week Mastercard made sure the pattern holds.

Mastercard confirmed a deal with the startup Alchemy that lets AI agents buy things on your behalf. You link a card, set a spending cap and a list of approved merchants, and the agent orders the food or books the trip without checking in each time. It is a clean product and an obvious one. The feature is easy to picture. What it locks in underneath is the part worth reading closely.

The disruption that leaves the toll booth standing

Look at what the deal leaves untouched.

Alchemy had already built its AgentCard product into Visa's Intelligent Commerce back in June. Mastercard makes two, which puts most of the cards in your wallet behind a single rail for agent payments. American Express has its own version, and PayPal and Stripe are building rival frameworks. The models will turn over, the agent wrappers will turn over, and Alchemy may well be right that the checkout button is dying.

Every one of those agent purchases still ends at one of the same four networks, clipping the same fee it has always clipped.

Read that again. The technology sold as the thing that finally gets around the middleman instead turns the middleman into the one piece nobody can pull out. The models are commodities in waiting. The agents are wrappers. The networks stay. Markets read it correctly and shrugged: Mastercard and Visa shares barely moved on the news.

So the toll booth survives. Fine. That was always the boring half of the story.

"Engineered trust" is a claim, not a comfort

The sharper half showed up on LinkedIn.

A Mastercard executive has a post going around to promote a talk at a risk conference in Arizona. It opens on the reasonable question of the moment, what happens when AI starts acting on our behalf, and lands on a line worth reading twice. As commerce moves from human clicks to autonomous agents, the post says, trust must be "engineered into every machine decision," and security has to keep pace with autonomy.

It reads like a thoughtful sentiment. It is a land grab.

In agentic commerce, whoever engineers the trust owns the definition of intent. And whoever owns the definition of intent owns the thing that matters when money moves: liability.

Who gets to define intent decides who eats the loss

Here is how it works.

The deal runs on Mastercard Agent Pay, and its centerpiece is a framework Mastercard calls Verifiable Intent, which binds your identity, your instructions, and the outcome of a transaction into one tamper-resistant record. Mastercard presents it as consumer protection, and in part it is. The agent pays with one-time-use tokens instead of your real card number, so the card underneath stays walled off. Good.

Now ask what that sealed record of your authorization is for.

American consumer credit protection runs on one bright line. Was the charge authorized, or was it not. Unauthorized charges are the bank's problem, and the chargeback is your weapon. Agentic payments smear that line into a gradient. You did not approve the specific purchase. You approved an agent, inside some limits, to make purchases like it.

So when the agent books the wrong nonrefundable flight, or a merchant games the comparison the agent is running, who absorbs the loss?

Verifiable Intent is built to answer that in the network's favor. Here is proof the customer consented. The safeguard handed to you doubles as the instrument that shrinks what you can dispute later. This is the host of questions about customer protections and fraud risks the Wall Street Journal waved at. Except it is less an open question than a design decision, and it is being made for you by the company writing the standard.

The referee would also like to play

Which is what the LinkedIn post was telling us all along.

When a Mastercard executive says trust must be engineered into every machine decision, the word doing the work is not trust. It is every. The pitch is to sit underneath each autonomous purchase and rule on whether it counts. A referee volunteering to also play, across a large share of the world's transactions, before anyone has agreed on the rules of the game.

None of this makes agentic payments a bad deal for the people using them. Spending limits, approved-merchant lists, and tokenized card numbers beat handing a bot your live credentials, and the convenience is real. If you want an agent to reorder the coffee, this is a better way to let it.

But do not mistake what happened this week. It is not that AI learned to shop. It is that the hardest industry in the whole stack to disrupt found a way to make the disruption depend on it, and started writing the definition of "you authorized this" while everyone was still watching the demo.

The agents are new. The toll booth is old. And it just added a rule that says whatever the agent does with your money, there is now a receipt proving you asked for it.