The AI Layer Took the Glamour. We Kept the Chargeback.
Anthropic's commerce blueprint ships everything except the part that decides who pays when an agent gets it wrong
Anthropic's commerce blueprint ships everything except the part that decides who pays when an agent gets it wrong.
I read the release twice looking for the payment layer. The second time, I realized its absence was the point.
What shipped
On 2 September, Anthropic released a blueprint that gives retailers prebuilt code for two commerce agents. A shopping agent searches a catalog, compares products and builds a cart. A merchant agent handles inventory, pricing and marketing on the other side of the store. The merchant agent's recommendations require human approval before anything changes. Reference implementations cover retail, travel, telecom and ticketing, and Shopify and Priceline are named early users. Visa, Mastercard and Accenture will help carry it into their merchant networks and client bases. Anthropic's pitch is speed: an engineering team with a working commerce agent in days rather than months. Anthropic Debuts Commerce Agent Blueprint With Visa and Mastercard +3
The results Anthropic cited are real but narrow. For one partner, cart size rose roughly 30–35%, and customers were about 60% more likely to complete a purchase. One partner, company-reported. Price those numbers accordingly. investing
And then the part that matters. The blueprint includes no checkout protocol, no payment infrastructure and no advertising mechanism; those stay with retailers and their chosen partners. The shopping agent assembles the cart and hands it off. Authorization happens somewhere else, on rails Anthropic chose not to touch. fintech
The omission is the thesis
This isn't a gap. It's a map of where value and risk actually sit.
Discovery, comparison and cart-building went to the model layer — and they belong there. They were never defensible. Any sufficiently good model can read a catalog.
What didn't move an inch: authorization, tokenization, dispute rights, liability. That's our layer. It's the layer that decides, transaction by transaction, who is left holding the loss. And it was built on an assumption that is quietly expiring: that a human decided to buy.
Where the rules crack
Three places, concretely.
Who initiated it. Card rules sort card-not-present transactions into two families: the cardholder initiated it, or the merchant initiated it under a prior agreement with the cardholder. An agent purchase fits neither cleanly. The cardholder didn't press the button. The merchant didn't start the transaction. Our systems will force it into one of the two boxes anyway — and whichever box we pick, the dispute rights follow.
What evidence wins. Today, a merchant fighting a card-absent fraud dispute on Visa can use Compelling Evidence 3.0: show prior undisputed transactions from the same cardholder that share identifiers like the IP address or the device. That's the fingerprint of a real person on a real phone. An agent running in the cloud doesn't transact from the cardholder's phone. The identifiers that win disputes today will increasingly point at a data center.
Who was authenticated. Liability shifts to the issuer when the cardholder is authenticated. But when an agent acts hours after the human last touched anything, what exactly was authenticated — the person, the agent, or the permission the person once gave the agent? The networks and the protocol builders are working on it: Visa Intelligent Commerce, Mastercard's agentic tokens, and protocols like Google's AP2 that record signed mandates of what a user authorized. That's genuine progress. But evidence is only worth what the dispute rules say it's worth. And those rules haven't been rewritten.
Every dispute team in this industry is about to meet a new sentence: "I didn't buy that. My agent did." Is that fraud? First-party misuse? Agent error? Merchant misrepresentation? Each answer maps to a different reason code — and a different party paying.
From a Panamá desk
In a dollarized market, our cardholders buy from U.S. merchants as easily as from a store down the street. So picture the first agent dispute here: an issuer in Panamá, a merchant and acquirer in the United States, an agent platform somewhere else, and dispute rules written for none of the three.
From where I sit, a single chargeback rule change can reshape a quarter of operational work. An undefined category of transaction isn't a thought experiment. It's a backlog waiting to happen.
Issuers and processors in smaller markets don't write network rules. We inherit them. Which is exactly why we need to be in the conversation before they're written, not after.
My judgment
The AI layer took the glamour and left us the liability.
Discovery was never the moat. The moat — and the risk — is who eats the chargeback. The industry has started shipping ways to record what an agent was allowed to do. It has not shipped agreement on who pays when the agent does something else.
That agreement won't come from model companies. It will come from the people who sign off on dispute rules. So that's who I want to hear from.
Three questions for the people who own this
- When an agent buys, who is the cardholder of record in your systems — and can your authorization flow even express that an agent was involved?
- Which liability shift do you apply, and who at your institution actually has the authority to decide it?
- If agent-initiated volume hit 5% of your card-not-present traffic tomorrow, does your dispute flow bend or break?
Reply to this email with even one line. The sharpest answers will shape the next issue.
Sources
- PYMNTS, 3 Sep 2026 — Anthropic debuts its commerce agent blueprint and deliberately leaves out the wallet. https://www.pymnts.com/news/artificial-intelligence/2026/anthropic-built-the-shopping-brain-and-skipped-the-wallet/
- Finextra, 4 Sep 2026 — Anthropic launches AI commerce agents with Visa and Mastercard. https://www.finextra.com/newsarticle/48346/anthropic-launches-ai-commerce-agents-with-visa-and-mastercard
- Reuters, via Investing.com, 2 Sep 2026 — Anthropic launches AI agent blueprints for retailers ahead of the holiday season. https://www.investing.com/news/economy-news/anthropic-launches-ai-agent-blueprints-for-retailers-ahead-of-holiday-shopping-season-4886525