AI ecommerce agents: products, orders and actions

Which shop problem should the agent solve?

Choose one journey where customers need context: a technical variant, compatible accessories or an allowed order change. An agent for the entire shop is difficult to evaluate without a predefined correct outcome.

A sales assistant succeeds with a suitable product and understandable reasoning. A service agent succeeds with a correctly completed action. They may share a chat window but require different data and permissions. Public product information does not need the same verification as personal customer data.

Start with a category whose common questions and return reasons you understand. Good automation can prevent unsuitable orders. A conversion increase followed by more returns and support costs may provide no business benefit.

Why does the correct variant matter more than a polished description?

Customers buy a specific size, colour, configuration or compatible component. A model can write a convincing recommendation without evidence that two items work together. Product relationships must therefore be supported by data.

A shop owner on Reddit in July 2026 seeks help with a large range and widely varying prices. Another thread about orders originating from ChatGPT discusses unsuitable parts recommendations. These unverified experiences do not establish error rates, but suggest useful test questions.

Keep product and variant identifiers with every recommendation. “Black adapter” is insufficient if three versions exist. Missing specifications should trigger a follow-up question. The operator must be able to define mandatory recommendation fields.

What changed in agentic commerce in 2026?

Shopify’s June Spring ’26 announcement connects its structured Catalog and UCP shopping protocol with AI channels. It reports twice the conversion rate for Catalog-based search compared with scraped data, 3.2% of Omnilux’s March revenue from AI channels and and 20-fold year-on-year growth of that channel at Cozy Earth. These are different figures from the vendor and individual brands. They do not predict the benefit of your chat. Shopify announcement, 17 June 2026.

“Merchants can now ask Sidekick real questions.” — Anne Prins, VP of Product Partnerships at Klaviyo, shortened quotation from Shopify’s announcement.

The direction matters: AI connects to operational data and tools. It does not mean every platform offers the same interfaces. Before custom development, check what the shop platform already solves and what requires a separate connection.

Which sources need connecting?

At least the catalogue and current business information must be connected. Personal service adds orders, delivery status and customer verification. Sources change at different rates, so a one-off upload into one document is unsuitable.

For technical products, link information to the exact model and production period. General category copy must not override manufacturer specifications. If sources conflict, record that conflict and offer a verified alternative or further contact.

How should carts and order changes work?

The agent should use the existing checkout mechanism to validate price, quantity, variant, discount and delivery. The model must not independently assemble a binding total from numbers in the conversation.

Before completion, customers see specific items and terms. If a price has changed, the next step uses the current state and explains the difference. Payment processing remains in the designated payment environment.

For existing-order changes, check again whether the action is still possible. Changing an address before or after carrier handover involves different processes. The agent needs the connected system’s real capabilities, not just generic “contact support” guidance.

What goes wrong in real shop implementations?

Common issues are inconsistent specifications, delayed stock updates and unclear ownership of product data. Tests often reveal multiple names for the same property, incomplete variant dimensions or documentation for a different model.

In the April 2026 Shopify discussion about AI Toolkit participants distinguish easy text tasks from variant and formatting problems. These experiences should be checked against your own data. Standardising specifications alone can improve search and filters, regardless of AI.

Integrations need their own error monitoring. If the catalogue stops responding, the agent must not confidently recommend from memory. It can explain the limitation, answer general questions and offer to continue when service returns.

Which pilot can demonstrate business value?

Use one category, a few clear scenarios and comparison with the current process. Test recommendation accuracy, customer time, completed purchases and returns. For service, measure repeat contacts and success of specific actions.

Illustrative calculation: an agent influences 100 orders with a CZK 300 contribution margin each. That is not automatically CZK 30,000 in new value. Establish how many orders were genuinely additional, whether they replaced normal purchases and their servicing costs including returns.

If the pilot exposes inadequate data, fixing it may be the next step rather than changing models. For broader project prioritisation, see AI for ecommerce. We then design a specific agent where it has a demonstrable advantage.