On September 2, 2026, Anthropic open-sourced a retail commerce agent blueprint built on Claude. The release includes two reference implementations. A shopping agent that searches a retailer’s catalog, compares products, builds a cart, and answers customer service questions within a single conversation. And a merchant agent that analyzes sales performance, monitors inventory, suggests pricing, and drafts marketing campaigns based on the retailer’s own data. Partners including Shopify, Priceline, Accenture, Mastercard, and Visa are already working with the blueprint. Anthropic reports carts growing up to 35 percent. Shoppers are also 60 percent more likely to complete a purchase for retailers using Claude shopping agents.
The most important sentence in the announcement is not about what the agent does. It is about what it deliberately does not do. The blueprint contains no payment protocol, no checkout layer, and no transaction routing through Anthropic. The shopping agent builds the cart and hands it to the retailer’s own checkout. Full stop.
That design decision is not a technical limitation. It is a design philosophy. And it is the most accurate read of where consumer trust actually sits in retail AI in 2026. Anthropic built the shopping brain and skipped the wallet. Here is why that is the right call, and what it means for every retailer building an AI commerce strategy before the holiday season.
The Number That Explains the Design Decision
A Gartner survey found that only 11 percent of consumers are willing to let AI make buying decisions for them. That is not a temporary hesitation. It is a structural condition of where trust sits in the AI commerce relationship right now. Interestingly, the same research shows consumers are eager to use AI for product discovery and comparison. However, they want to retain control of the actual purchase decision.
This aligns precisely with what I described in the Year One agentic commerce analysis. The dominant pattern across 2025 and 2026 is agent-assisted discovery followed by checkout on the retailer’s own site. Walmart found conversion three times higher when redirecting AI-referred shoppers to its own checkout than when completing the transaction inside the chat. In fact, OpenAI ended Instant Checkout by March 2026. The consumer behavior data from Year One and the Gartner trust research say the same thing from different angles. The agent is trusted to help. It is not yet trusted to decide.
Consequently, Anthropic designed for the 89 percent who want AI assistance during discovery and then check out on the retailer’s own platform. Not for the 11 percent who are ready for fully autonomous buying. That sequencing is correct for where the market is today. Moreover, it is the sequencing that produces the best conversion outcomes, as the Walmart data confirms.
What the Blueprint Actually Gives Retailers
The Shopping Agent: Discovery and Handoff
The Claude shopping agent sits inside a retailer’s own app or website. The agent searches the retailer’s catalog, retains customer preferences within the session, and processes multi-item requests. Within the conversation, it presents products and comparisons, builds a cart, and hands the completed cart to the retailer’s existing checkout. Order tracking, returns, exchanges, and refund policy questions are also handled in the same conversation.
The architecture is deliberate. Claude sits on top of the retailer’s existing infrastructure as a reasoning layer. Importantly, it does not compete with the retailer for the customer relationship or the payment flow. Shopify’s implementation shows the handoff clearly. The agent finds products and builds the cart. Then it passes the shopper to the store’s existing checkout, where payment stays. Anthropic does not route or see the transaction at any point.
In other words, this addresses the data ownership concern I described in the AI retail discovery and transaction analysis. The retailer who deploys a Claude shopping agent retains the transaction data, the customer relationship, and the loyalty program linkage. The agent generates intelligence. The retailer captures the value.
The Merchant Agent: The Underreported Story
Most coverage of the Anthropic blueprint focused on the shopping agent. The merchant agent is equally significant and almost entirely underreported. It targets internal retail operators, not customers. It answers questions about sales performance, monitors inventory levels, suggests pricing changes, and drafts marketing campaigns based on the merchant’s own data. All outputs go to a human for approval before anyone takes action.
Furthermore, this connects directly to the inventory accuracy challenge I described in the agentic commerce inventory accuracy analysis. The agent that helps store operators monitor inventory, identify phantom stock, and adjust pricing addresses the exact operational problem that causes agentic commerce to fail at the customer-facing layer. Indeed, the merchant agent and the shopping agent are not separate products. They are two sides of the same operational loop. The merchant agent improves the data quality that makes the shopping agent reliable.
The Three Guardrails That Make This Deployable
The blueprint includes guardrails that Anthropic built into the reference implementation. They matter because they address the specific trust concerns that the Gartner data documents.
Catalog Constraint: The Agent Cannot Invent Products
The shopping agent is limited to the retailer’s actual catalog data. It cannot recommend products that do not exist in the retailer’s inventory or cite prices that differ from actual pricing. This addresses the hallucination risk that makes AI commerce untrustworthy. Specifically, the agent cannot recommend a product that does not exist or quote a price that does not match. As I described in the Sephora AI commerce analysis, the quality of the recommendation is directly determined by the quality of the underlying product data. The guardrail enforces that dependency by design.
No Manipulative Upsell: The Agent Recommends, Not Pressures
The blueprint specifically prevents manipulative upselling behavior. The shopping agent recommends based on the customer’s stated preferences and catalog match, not on margin optimization or promotional pressure. This is a trust-preserving design decision. That 11 percent willing to let AI make buying decisions will not grow if AI agents optimize for retailer margin rather than customer intent. That is a trust-destroying dynamic at scale. The guardrail addresses that risk before it becomes the dominant pattern.
Human Approval for Merchant Actions: The Agent Suggests, Not Acts
Every merchant agent output requires human approval before anyone takes action. This mirrors the governance framework I described in the eTail Boston AI adoption analysis. The retailers who get AI deployment right define clear decision ownership. Which calls belong to the AI? Which to the human? What happens when they disagree? The Anthropic blueprint builds that governance into the reference implementation rather than leaving it to the retailer to design after go-live.
What This Means for Retailers Building Before the Holiday Season
As a result, the Anthropic blueprint is open-source, free, and deployable through Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI. Specifically, it requires the retailer to connect their own catalog, cart, and checkout infrastructure. Anthropic will not maintain it as a product. Retailers who deploy it own the implementation and the ongoing maintenance.
That ownership requirement is the most important detail for retailers evaluating the blueprint before the holiday season. Ultimately, the blueprint provides the reasoning layer. The retailer provides the data layer underneath it. As I described in the agentic commerce inventory accuracy analysis, reliable agent behavior requires product data at 95 percent or higher fill rates on core attributes. A catalog with incomplete attribute data, inaccurate on-hand numbers, or stale pricing will produce a shopping agent that builds carts the retailer cannot fulfill. The blueprint does not fix the data layer. It exposes it.
What This Means for LatAm Retailers
The open-source nature of the Anthropic blueprint is particularly relevant for LatAm retailers. The Apache 2.0 license means any retailer can deploy the reference implementation without licensing fees. Notably, Heritage Grocers Group already deployed the Spanish-language version of a similar shopping assistant. That precedent matters for LatAm markets where Spanish-language AI commerce infrastructure has lagged behind English-language deployments.
Additionally, the merchant agent’s inventory and pricing functions address a specific LatAm challenge: catalog management across diverse store formats and regional assortment variations. A merchant agent that surfaces inventory discrepancies, flags pricing inconsistencies, and drafts campaigns in the operator’s own language reduces manual overhead significantly. That reduction matters in LatAm formats where technology teams are smaller and AI commerce has historically required more manual coordination.
The Question the Blueprint Forces Every Retailer to Answer
What Your Data Layer Can Actually Support
The Anthropic blueprint is the most accessible AI commerce infrastructure ever available to retailers. It is free, deployable through existing cloud relationships, and backed by early results no other open-source retail AI project has produced. However, it surfaces a question no blueprint can answer for you. Is your product catalog ready? Is your inventory data accurate? Can your checkout infrastructure convert the AI-referred click-through fast enough?
By contrast, an agent connected to a catalog with 60 percent attribute fill rates will not produce 35 percent larger carts. It will produce frustrated shoppers and failed recommendations. The blueprint is the reasoning layer. The retailer owns the data layer. That division of responsibility is not a gap in the blueprint. It is the correct design. But it means the work that determines whether the blueprint produces results is the work that happens before deployment, not during it.
The Blueprint Is Ready. Is the Store?
Anthropic built the shopping brain and skipped the wallet. Gartner says 89 percent of consumers are not yet ready to let AI complete the purchase. Both are reading the same market correctly. Retailers who deploy the blueprint with clean data, accurate inventory, and a fast redirect will match the reference numbers. Others who deploy it on top of incomplete data, expecting the AI to compensate, will produce results that validate the skeptics. The blueprint is ready. Is the store?
If you are evaluating the Anthropic commerce agent blueprint or auditing your product data and inventory infrastructure before a holiday season deployment, connect with me here or reach me on LinkedIn. I am happy to walk through the framework we use across the U.S. and Latin America.
Adriana Rivas is a retail technology executive and AI strategist. She is the recipient of the Gold Stevie® Award, Thought Leader of the Year 2026, and recognized by Thinkers360 as the #7 Global Thought Leader in Retail. She is the author of How to Implement Self-Service Without Failing, with a Revised and Expanded Edition publishing September 21, 2026.