Claude Commerce Agents: Anthropic open-sources two AI agents for e-commerce (what to actually do with them)

Anthropic has released the Claude Commerce Agents for free and as open source: two AI agents built specifically for e-commerce. The first works on the customer side and helps shoppers find the right product; the second works on the management side and analyses sales, catalogue and inventory. It is worth reading carefully if you sell online, because it moves a piece of technology you used to rent as a subscription into something you can install and shape around your own store.

The sales agent, in practice

This is not the FAQ chatbot everyone knows. It genuinely queries the catalogue: it understands a request written in plain language — «I need a waterproof hiking shoe for my wife, size 39, under 120 euros» — filters real products, checks availability, offers two or three alternatives explaining the differences and handles objections. It is the shop assistant who, in a physical store, makes the difference between a customer who buys and one who walks out, and who simply doesn't exist online: in most stores the customer is left alone with a grid of 400 products and filters nobody uses.

The analytics agent, in practice

It looks at your data and answers questions that today take an afternoon in Excel: which products have high margin but low visibility, which SKUs haven't moved in 90 days and are tying up capital, which product pages are incomplete and convert worse, which items will go out of stock before the next reorder, which customers have stopped buying. You don't need a data analyst to ask those questions — you need someone to ask them regularly, and an agent does it every Monday morning without forgetting.

Why free and open source changes the maths

Conversational search on large stores is sold as a service today, with fees that start at a few hundred euros a month for a mid-sized catalogue and grow with traffic. An open source agent runs where you want, connects to your catalogue, and can be reshaped — tone, commercial rules, when to upsell and when not to — while you only pay model usage. For a store with normal volumes that means tens of euros a month, not hundreds in subscription. Your sales and customer data also stay where you decide, which matters for GDPR.

What it actually takes to go live, no illusions

The code is free, the project isn't. Four things are needed. One: a clean catalogue with product pages that carry real attributes (sizes, materials, compatibility) — poor product data means poor recommendations, because the agent advises on what it can read. Two: the connection to the store, whether Shopify, WooCommerce, Prestashop or a custom system, reading stock and prices in real time. Three: the commercial rules, meaning what the agent may and may not promise on discounts, shipping and returns. Four: supervision and measurement, because a sales agent should be judged on one number — the conversion rate of sessions where it intervenes versus sessions where it doesn't. A serious implementation on a mid-sized catalogue typically takes two to five weeks; in the projects I run the range is 3,000 to 12,000 euros depending on catalogue complexity and integrations, plus model usage.

When it pays off and when it doesn't

It pays off if you have a broad or technical catalogue where customers need guidance; if you get many pre-purchase questions by chat, email or WhatsApp; if you hold stock and want to stop discovering shortages too late. It pays off far less if you sell a handful of simple, self-explanatory products, if traffic is so low that any percentage gain is statistical noise, or if your catalogue data is a mess — in that case the first project is fixing product pages, not installing an agent.

The right way to start is small and measured: switch the sales agent on for a single category, the one where customers ask the most questions, for four weeks, A/B tested against traditional search. In parallel, run the analytics agent over the last twelve months of data and see what surfaces — the first report alone often justifies the project, because it quantifies dead capital nobody had measured. If the numbers hold, extend to the whole catalogue; if they don't, you spent little and learned a lot.

If you sell online and want to know whether these agents make sense for your catalogue, let's talk for 15 minutes: we look at the real store and the data you have, and I tell you which price bracket you fall into. If the return isn't there, I'll say so straight away.