The model is free now. Your catalogue is not.

On 2 September 2026 Anthropic published a repository under Apache-2.0 containing two commerce agents as a blueprint. Shopify, Visa, Mastercard, Accenture and Priceline are named as partners. For a week the coverage ran under the headline that AI now goes shopping.

That is not quite wrong, but it describes the wrong side. We looked at the blueprint and at Shopify's implementation, because we run a shop on Shopify ourselves and because the question of what this means in practice in the fourth quarter cannot be answered from a press release.

What was actually published

Two agents, and both run at the merchant, not at the customer.

The shopping agent is built into your own shop. It searches the shop's catalogue, assembles products, builds a cart and answers questions. Then it hands over to the regular checkout.

The merchant agent runs in the back office. It reads products, orders and inventory, proposes prices and campaigns, and stages changes for approval. Shopify's implementation solves this cleanly: the admin access token never reaches the model, changes are recorded and only executed after a human has approved them.

What the blueprint explicitly does NOT contain: payment processing. That stays in the existing checkout. And it contains no catalogue.

Just as important is what it is not. It is not a product. It is example code for developers. Shopify says so itself and points to a different in-house solution for the non-technical route. Anyone expecting to click this on in the admin area will not find it there.

The sentence that matters

The description of the guardrails states that prices and products are constrained to the actual catalogue data. And in Shopify's example code there is the sentence that sums the whole thing up: you point the agent at a shop domain, and it searches the live catalogue, builds a real cart, and answers from the shop's own policies and FAQs.

From the shop's own policies and FAQs.

That is the entire story, and it is less comfortable than the headline. The agent is not a salesperson who knows things. It is a reader, and it is exactly as good as what it is allowed to read. It can answer a question about delivery time, returns, suitability or accessories only if the answer is in the data. If it is not there, the agent says nothing, or says something unspecific, and the customer leaves.

The reasoning layer is being given away as of now. Openly licensed, running in days, deployable from four cloud platforms. What is not being given away is what it reaches into.

What this means for an online shop over the Christmas season

Here is the honest answer, and it is unspectacular.

Do not build this for this Christmas season. The blueprint is days old, it is developer example code, and it needs a team to integrate it, test it and keep it running. Retail usually freezes its systems in October or November. A conversational agent running under load for the first time in December, quoting a wrong price or a wrong delivery promise, is not a revenue lever. It is a customer service case.

What you can do instead pays off either way. Because the inputs are the same whether the agent belongs to you, to Shopify, or is the assistant on the other side where the customer asks the question. In all three cases the same catalogue is read, the same returns policy, the same FAQs.

Concretely, in the order in which it is worth doing:

1. Answer the questions that get asked before the purchase. Not the ones the marketing copy answers. Does this accessory fit? How long does delivery actually take, not in the best case? What is the difference between these two models? If those answers are nowhere, no agent can give them.

2. Put returns, shipping and warranty into text a machine can read. Not as an image, not as a PDF, not only in the footer.

3. Check your catalogue for completeness, not for beauty. We measured this on ourselves and were not proud of the result: out of more than thirty maintained product attributes, six appeared in the machine-readable markup. Customer reviews were visible on the page and entirely absent from the markup. The product description was cut off mid-word at eight hundred characters, and the part that fell away was precisely the one describing suitability.

4. Only then think about an agent. If you want one at that point, there is also a route for Shopify without your own development.

That is not a postponement, that is the order. An agent on a patchy catalogue does not make the gaps smaller. It makes them audible for the first time. Until now a customer traded a missing detail silently for another shop. From now on they ask, and they get an answer that either holds or does not.

What this means for a digital agency

For agencies the release is good news and bad news, and both concern the same position.

The bad: building a conversational agent has been a sellable project for two years. From now on it is considerably less so. When an open blueprint with reference implementations for four industries is available at no cost, and the vendor itself promises a team can get it running in days, integration becomes craft rather than a differentiator. Anyone who has been selling the ability to build a chatbot is now selling something the client can download.

The good: the value moves to both ends, and both are harder to copy than the middle.

Ahead of it lies the data work. Catalogue structure, attributes that actually support a purchase decision, policies in machine-readable form, answers to real pre-purchase questions. It is unglamorous, it is laborious, and it cannot be downloaded, because it looks different for every catalogue.

Behind it lies the safeguarding. What may the agent commit to? How is it verified that it does not invent prices? What does the approval process look like for the merchant agent when it proposes price changes? How do you measure whether it helps or harms? Shopify's implementation shows the direction by staging changes instead of executing them. But approval logic that fits a given company does not come out of the repository.

Any agency planning to keep selling build projects should expect the client to compare the price against an open blueprint. Any agency selling data quality and verifiability has just been handed an argument it did not have to invent: the maker of the model says itself that the agent does not get past the catalogue.

What this means for us

We do not write about this topic neutrally, and that belongs in the text. Klariton measures whether a shop's content is findable, complete and verifiable for machines. In that sense this release confirms our working premise, and that is exactly why we should be careful rather than enthusiastic in how we read it.

Three things genuinely change for us.

First, the input layer now has a name. We used to have to explain why product data matters for AI answers. Since 2 September that explanation ships with the model vendor, in its own architecture description.

Second, the object of measurement shifts. If an agent answers from policies and FAQs, then the question is no longer only whether a page is marked up, but whether the answer to a specific customer question exists at all. That is a different measurement from the one in common use today.

Third, and this is the uncomfortable side: an agent answering directly inside the shop is a closed room. What it says is invisible to any external checker. Our entire measurement method rests on querying from the outside what machines say about a shop. For the built-in agent that does not apply. We do not yet know how to solve this, and we consider it more honest to write that down than to pass over it.

What we do not know

Anthropic cites carts up to thirty-five percent larger and a sixty percent higher likelihood of completion. Those are the vendor's own figures, they are not independently verified, and they do not come from an environment comparable to a mid-sized German online shop. We list them here because they appear in every article on the subject, and we mark them as what they are.

Also open: how this behaves in operation when the catalogue is large, the variants numerous and the stock situation volatile, which is to say precisely during the Christmas season. Four days after the release there is no dependable experience on that, from anyone.

Sources and status: Anthropic, blueprint for commerce agents, published 2 September 2026 under Apache-2.0, with reference implementations for retail, travel, telecom and ticketing. Shopify's example implementation connects the shopping agent to a store via the catalogue interface and the merchant agent via the admin interface. Named partners: Shopify and Priceline with implementations, Accenture, Mastercard and Visa as integration partners. Status of this text: 6 September 2026.