Anthropic Shipped a Shopping Agent Blueprint and Left Checkout to Merchants

Anthropic shipped a shopping agent and refused to touch the checkout.

On 2 September 2026, Anthropic released an open blueprint for building commerce agents on Claude, with Accenture, Mastercard and Visa signed on to bring it to their clients and merchant networks.

Read the integration list closely: catalog, cart, checkout, customer preferences, order history. Then the line that matters: payment is left to you, whether that is your existing checkout or an agentic payments provider.

No payment protocol. No catalog. No ad layer. Nothing routes through Anthropic.

Two Models for Agentic Commerce

There are two broad ways AI agents are entering commerce right now.

The platform model puts a third-party agent between the merchant and the customer. This blueprint does the opposite. The agent lives inside the retailer’s own experience, and the merchant keeps the transaction.

Why That Is the Real Story

The reasoning layer is not where the margin sits

Models are converging fast. The durable position is the trust and authorization layer every agent has to call to get paid. That is exactly why Visa and Mastercard are partners in this announcement rather than competitors to it. Whoever authorizes the payment still owns the economics.

Merchants asked for this

Jack Forestell of Visa said merchants want more control over how AI engages their customers. Sherri Haymond of Mastercard framed it as merchants building their own agents. First-party is the demand signal, not only the vendor’s preference.

Bigger baskets meet unchanged risk rules

Anthropic reports carts up to 35% larger and shoppers 60% more likely to complete a purchase. Those are vendor figures, not an audited benchmark. But if they are directionally right, higher tickets and non-human session behavior will land on fraud engines tuned entirely on human browsing.

That is a real operational risk. Fraud models learn from signals like time on page, mouse movement, number of product views and basket composition. An agent-assembled cart looks nothing like a human session. Expect more false declines on good agent traffic until models are retrained, and plan for it. In my work, network tokenization delivered a 6% authorization lift on card-on-file, but none of that decisioning logic was built for agent-assembled carts. I wrote about the cost of declining good customers in Fraud Declines vs. No-Fraud Declines.

Owning the cart means owning the data

A merchant-hosted agent keeps preference and intent signals inside the merchant’s stack. A platform agent exports them. Over time, that data decides who can personalize offers, who can price loyalty and who can train the next model.

What Merchants Should Do Now

1. Decide where your agent lives. Build your own, plug into platform agents, or both. The answer can differ by category and channel.
2. Keep checkout yours. If the blueprint model spreads, your existing payment stack, tokens and fraud tools remain the system of record.
3. Flag agent traffic. Mark agent-initiated sessions in your data so fraud and authorization teams can measure them separately.
4. Retune risk rules. Review velocity limits, basket-size rules and device signals that assume a human shopper.
5. Talk to your issuers and networks. Agentic tokens and trusted-agent programs from the networks exist so issuers can recognize agent traffic instead of declining it. See Network Tokenization vs PCI Tokenization for the basics of how network tokens work.

My Read

The agent layer is being commoditized on purpose. An open blueprint that any merchant can use, with the networks on board and no payment layer of its own, pushes value away from the model and toward authorization, identity and the merchant relationship. That is good news for merchants who invest in their own payment stack.

If you run a merchant stack: build your own agent, or plug into whichever platform agent brings you traffic?

Commerce Agent Blueprint: Key Takeaways

• The commerce agent blueprint covers catalog, cart, checkout, preferences and order history, but leaves payment to the merchant.
• A merchant-hosted commerce agent keeps the transaction and the customer data with the retailer.
• Fraud and authorization rules need to be retuned before a commerce agent sends real volume.

FAQ

What did Anthropic release for commerce agents?

On 2 September 2026 Anthropic released an open blueprint for building commerce agents on Claude, covering catalog, cart, checkout, customer preferences and order history. Accenture, Mastercard and Visa signed on to bring it to their clients and merchant networks.

Does the Anthropic blueprint process payments?

No. Payment is left to the merchant, either through its existing checkout or an agentic payments provider. There is no Anthropic payment protocol, catalog or ad layer, and nothing routes through Anthropic.

What is the difference between a merchant-hosted agent and a platform agent?

A merchant-hosted agent lives inside the retailer’s own app or site, so the merchant keeps the transaction and the customer data. A platform agent sits between the merchant and the customer and tends to pull checkout, and the data, toward the assistant.

A shorter version of this analysis was first published on LinkedIn on September 5, 2026.

Related Reading

• What Is a Risk Decisioning Platform? Workflow, AI, and Design Principles
• Authorization Rate Optimization: A Payments PM’s Complete Guide
• Fraud Declines vs. No-Fraud Declines
• Network Tokenization vs PCI Tokenization: What’s the Difference?