Claude for Commerce Didn't Replace Composable Architecture. It Proved the Case for It.

On September 2, Anthropic released Claude for Commerce, an open-sourced pair of reference agents—one built to shop for a retailer's customers, one built to run a merchant's back office—that any retailer can deploy for the cost of the tokens. The blueprint ships under an Apache 2.0 license, with eight working reference implementations spanning retail, travel, telecom, and entertainment, and early partners that already include Shopify, Visa, Mastercard, and Accenture. Anthropic says retailers running Claude-built shopping agents have seen shoppers add more to their carts and become 60% more likely to complete a purchase. It’s a vendor-sourced claim, not an independent finding, but it’s notable all the same.

Much of the coverage this week reads it as a threat: another AI lab moving into commerce, one more reason retailers might not need their existing platform stack. But that reading skims past a pretty important part of the announcement.

To make the shopper agent work at all, Anthropic had to build a product ontology first: a master catalog of "hero SKUs," a system for canonicalizing offers and listings against them, a taxonomy, a datafeed, and product-crawl capability to keep it current. That's not a feature of the agent. That's the infrastructure the agent sits on top of, and Anthropic had to construct it from scratch because most retailers don't hand over anything close to it.

This shouldn't be surprising if you look at how these agents actually fail. Product content written for a human shopper—marketing copy, lifestyle photography, a persuasive paragraph—doesn't give an agent anything to reason over. Agents resolve against machine-readable fields: material, dimensions, compatibility, certification. Structured data isn't a nice-to-have layer on top of good content; it's the layer the agent actually reads. That's borne out beyond Anthropic's own build: 71% of the pages ChatGPT cites, and 65% of the pages Google's AI Mode cites, carry structured markup. It’s not a guarantee of citation, but a clear floor for participation.

Which is the part of this story composable and MACH buyers should recognize immediately, because it's the same floor their architecture already clears. Decoupling product, content, and commerce data behind APIs so any front end can consume a consistent, canonical version of "what do we sell, in what variants, at what price, in stock where" is the whole discipline. A retailer running a real composable stack has already done, as ordinary infrastructure work, the thing Anthropic had to reconstruct as a prerequisite for its own product. Composable didn't anticipate agentic commerce. It just happened to already require the one thing agentic commerce turns out to depend on.

That's a real head start, and it's a widening one: Gartner's tracking shows 67% of enterprise retailers now running some form of headless or decoupled architecture, up from 38% in 2022, with a forecast that 60% of new commerce builds will align to MACH by 2027. The retailers on the other side of that number are the ones who will feel this most.

A monolithic platform doesn't just lack an agent, it typically lacks the clean, canonical product model an agent needs to function at all— inconsistent SKUs across channels, content locked in a template layer, no real taxonomy underneath the merchandising. Those retailers now face the same ontology-building project Anthropic just completed, except they'll do it retroactively, under competitive pressure, instead of as planned infrastructure work. Checkout.com's own research found 72% of merchants believe consumers will adopt agent-led shopping faster than merchants themselves will be ready for it. The gap isn't agent selection, it's the data underneath.

Which is the practical point for anyone weighing what to do this quarter. The decision in front of most retailers isn't which agent or which protocol to bet on (Claude for Commerce, OpenAI's commerce push, whatever ships next). Anthropic's own build shows that question is close to secondary. The decision that actually matters is whether your product and content data is structured, canonical, and API-addressable enough for any agent to read it correctly. That's an audit you can run today, independent of which vendor wins the protocol war, and it's exactly the discipline composable architecture already forces.

None of this means composable commerce "wins" because of AI, and it would be a mistake to read it that way. The narrower and more useful understanding, though, is that the more commerce gets mediated by agents, the more a retailer's underlying data model determines whether those agents work at all. Anthropic didn't set out to make that argument. It just had to solve the problem before it could ship anything. And in solving it, it showed its work.