For most of the past year, agentic commerce has been two separate demos. An agent that could find the product. A different agent, in a different press release, that could in theory pay for it. The two halves rarely met in production. This week they did.

On June 3, Amazon began offering the technology behind its own shopping agent to other retailers through AWS, starting with Kate Spade New York. The product, an Agentic Shopping Assistant, lets a brand drop a buying agent onto its own storefront, Digital Commerce 360 reported. A day earlier, ING, Worldline and Mastercard completed what they call Europe's first end-to-end agentic payment transaction, according to Finextra. An agent did not just choose the cart. It cleared it.

The interesting thing is not that an agent can shop, or that an agent can pay. It is that, as of this week, the same flow can do both.

The loop, and why closing it changes the question

We have tracked these halves separately because they lived in separate places. We wrote about the buy side when agentic commerce first went live with three proof points. We wrote about the pay side, and its problems, in our piece on the agentic payments identity crisis. Discovery and settlement, two ends of the same transaction, kept getting built by different teams with different assumptions.

Amazon's move matters because it is not a demo. It is distribution. A retailer that switches on the Agentic Shopping Assistant is handing the first moment of the purchase, the part where a human used to browse, to a model. Cotopaxi, the outdoor brand, is already restructuring its catalog data so agents can read it cleanly, Digital Commerce 360 noted. When the buyer is software, your product page is an API whether you designed it that way or not.

The ING transaction matters for the opposite reason. It is small, controlled, and European, and it answers the question the buy side keeps dodging: when the agent decides, who actually moves the money, and under whose rules. A card network, a processor and a bank ran that flow together. That is the boring, load-bearing part. It is also the part that has to work before any of this is real.

Name the gap: a closed loop still has a hole in the middle

Here is the part that should worry you. Closing the loop does not close the trust gap. It moves it.

Think about a purchase through the MM Trust Layer Model, our frame for the three kinds of trust an agent transaction needs: trust in discovery (is this the right product, from a real seller), trust in authorization (did the human actually agree to this), and trust in settlement (can the money move and be reversed if it shouldn't have). Amazon just industrialized the first layer. ING, Worldline and Mastercard just proved the third. The middle layer, authorization, is the one nobody has finished.

That gap is not theoretical. We wrote a whole piece on the missing agentic dispute layer because the chargeback system has no code for "the agent bought the wrong thing." A human who fat-fingers an order owns the mistake. An agent that misreads a prompt and books the wrong flight does not. Someone eats that cost, and right now the contracts do not say who.

As Simon Taylor has argued in Fintech Brainfood, the AI commerce stack is turning out to be vertical rather than horizontal, each layer owned by a different player rather than one open rail. A closed loop built from vertically-integrated parts is convenient. It is also a place where liability can quietly fall between the parties, because each one assumes the next one caught it.

What merchants and issuers do now

If you run a store, the discovery shift is the immediate one. Shoppers are starting their journey by asking an assistant which product fits, not by typing into your search bar, Total Retail put it bluntly: your next customer might be an AI, and the question is whether your store is ready to be quoted by one. Clean, machine-readable product data is no longer an SEO nicety. It is shelf space.

If you sit on the payments side, the work is less visible and more urgent. The NMI CEO told PYMNTS that AI now shapes how the company thinks about acquisition, because consumers already pick among cards, wallets, account-to-account transfers and real-time rails by context rather than habit. Add an agent making that choice on their behalf and the routing decision moves up a layer, away from the cardholder and into the model. Whoever the agent trusts to authorize and settle gets the volume.

What to watch

Watch whether the authorization layer gets a standard or stays proprietary. Amazon's assistant authorizes inside Amazon's world. The ING flow authorizes inside a bank consortium's world. Those two worlds have to agree on what a valid agent instruction looks like, or every closed loop becomes its own walled garden with its own liability rules.

And watch the first dispute. Not the first transaction, the first one that goes wrong, gets contested, and forces someone to write down who pays. That is when we will learn whether this loop is actually closed or just looks closed from the outside.

The infrastructure is ahead of the demand again. That is either the smart early bet or a rail laid for traffic that has not arrived. This week it got one stop closer to an answer.

Sources

If the agent picks the product and the network clears the payment, who answers for the purchase in between?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.