Anthropic launched a commerce-agent blueprint on September 2, 2026, with a practical claim attached: retailers running shopping agents on Claude have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase.

That claim raises an operating question. When an AI agent helps someone search, compare, decide, and buy, can the company’s systems give it enough clean information to complete the work?

TL;DR

Shopping agents are moving commerce from assisted browsing toward delegated buying. Anthropic’s blueprint gives teams reference patterns for shopping and merchant agents in retail, travel, telecom, and ticketing. Stord’s 2026 e-commerce report says consumer use of generative AI for online shopping rose from 38% in 2024 to 51% in 2025, with 17% of shoppers using AI regularly. Leaders can start by making product data, inventory, policies, trust signals, and exception paths agent-ready.

Map the workflow around checkout

The easy read is conversion. Bigger carts. More completed purchases. Better assisted shopping.

That is real, but too narrow.

In Anthropic’s description, enterprise customers including Shopify and Priceline are using agents so consumers can search in plain language, find options, compare them, and buy. The new blueprint includes reference implementations for a shopping agent and a merchant agent, plus guardrails and patterns meant to get an engineering team running in days.

That matters because the purchase is only the visible tip of the workflow.

Before a customer buys, the agent may need product attributes, availability, delivery windows, return rules, promotion logic, warranty terms, substitutions, reviews, and payment boundaries. After the purchase, it may need order status, support paths, cancellation rules, and escalation handoffs. Scattered records and policies written for manual interpretation leave the agent with the same fragmentation your team already handles.

One retailer can let an agent compare, configure, and complete a transaction with confidence. Another can only answer generic questions and then hand the customer back to a normal checkout flow. The customer experiences that difference as friction in the purchase.

Agent-mediated buying rewards operational clarity

Stord’s 2026 State of AI in E-Commerce Report says 51% of consumers used generative AI for online shopping in 2025, up from 38% in 2024. It also estimates that 17% of consumers regularly use AI for shopping, roughly 45 to 50 million U.S. consumers.

Those figures describe consumer use of AI for shopping; the share of purchases completed autonomously remains a separate question.

The important shift is that AI changes where comparison happens. A shopper used to compare by opening tabs, reading descriptions, checking reviews, and deciding which tradeoffs mattered. Now they can ask an assistant to narrow the field before they ever see your site. If your product information is vague, incomplete, contradictory, or trapped in a layout the agent cannot use, you may lose before the customer arrives.

Making that information usable requires work across the operating model as well as the website.

Can the agent tell what is actually in stock? Can it distinguish a final sale item from a normal return? Can it explain the difference between two service tiers without inventing a policy? Can it route a complex order to a human before money moves? Can it avoid recommending an unavailable bundle? Can it say no when the answer should be no?

Commerce agents depend on clear processes to answer those operational questions reliably.

Start with a focused readiness check

A lot of teams will overreact to this category. They will hear “agentic commerce” and assume the project requires a full platform rebuild, a custom agent team, and months of architecture work.

Sometimes it will. Most of the time, the first pass is much smaller.

Start by mapping the buyer questions an agent would need to answer before recommending you. What is this product for? Who should not buy it? What does it cost? What changes the price? What is available now? What are the return limits? What proof exists that it works? What happens when the order is unusual? Where should the agent stop and involve a person?

Then audit where those answers live.

If pricing is in one system, exclusions are in another, inventory is delayed, support policies are written in internal shorthand, and reviews are detached from product attributes, a commerce agent will inherit that fragmentation. Check a friendly answer against those records before trusting it.

The practical work is to make the business legible. Clean product records. Clear policies. Current inventory. Structured proof. Named escalation paths. Human approval for edge cases. A written definition of what the agent is allowed to do, what it may recommend, and where it must stop.

That preparation gives the agent reliable information and usable boundaries.

The gap will look like convenience

Customers should find a prepared company easier to buy from through a shopping agent.

Their products will be easier to compare. Their policies will be easier to understand. Their checkout flow will tolerate delegation. Their support paths will be ready when the agent cannot finish alone. Their teams will know who owns the agent’s behavior because they assigned responsibility for it inside the business.

Poor preparation can quietly cost opportunities. An agent may skip a company because another answered better, or a shopper may abandon a cart because one exception required too much work. Investigate product data, policy drift, and ownership of the handoff when diagnosing those failures.

Anthropic’s 35% larger-cart and 60% completion claims are early numbers from a specific provider ecosystem. They should not be treated as universal promises. But they do point to the direction of travel: buying is becoming something agents can participate in directly.

Agent participation makes operational readiness part of the buying experience.

Before launching a shopping agent, test whether it could safely understand, explain, and complete the buying workflow you already have. Use any failures to prioritize cleanup before the build.