Agentic Commerce: What Product Data AI Agents Need and Where It Comes From

For agentic commerce, what counts is product data that an AI agent can read without having to ask. This includes attributes as structured values, separate units and variants, up-to-date prices and texts in every sales language.

AI tools that enrich product data in the feed quickly reveal where information is missing. To keep these gaps closed for good, product data management transfers the added values into the PIM system. Customer-specific prices, stock levels and delivery times originate in the ERP and change constantly. An agent ordering on behalf of a customer therefore needs a connection to the ERP, not just a catalog.

In Germany, AI assistants have so far mainly been used to search and compare, while purchases are usually still made in the supplier’s online shop (as of October 2026). But that is set to change.

It all starts with the question of which system manages which information, because feeds and interfaces for the assistants build on exactly that.

At the end of September 2026, Commerce, the company behind BigCommerce and Feedonomics, introduced two AI tools that fill in missing information in the shop and the product feed. According to the vendor, this should help assistants such as ChatGPT, Gemini and Copilot find and recommend the products more easily. The announcement shows how important the feed is becoming as a route to the assistants, and agentic commerce was already one of the key topics in our review of BigCommerce Commerce Live 2026 in London.

Tools like these reveal where information is missing and quickly fill the gaps for recommendations. To keep the gaps closed for good, the added values belong where the product data is maintained. After all, the feed, i.e. the regularly generated file through which product data reaches marketplaces, price comparison sites and assistants, only passes on what is stored in the systems behind it.

Attributes, variants and languages are usually managed in the PIM system. Customer-specific prices, stock levels and delivery times, on the other hand, originate in the ERP and change there constantly, which is why the feed reads them from there in real time. For companies, agentic commerce is therefore first and foremost a question of which system is responsible for which information.

An AI agent breaks a task down into requirements, searches for suitable items, compares them and checks price and availability before it recommends or orders anything. For each of these steps, it needs values it can interpret without having to ask.

A person ordering in an online shop reads “suitable for all common pipe sizes” and understands that it is worth asking. The agent, by contrast, looks for a nominal diameter attribute with a value such as “DN 25”, and if it is missing, it skips the item or guesses. Product data therefore only becomes readable for an agent when five things are right:

• attributes as structured values instead of in running text
• unique identifiers such as GTIN and manufacturer part number, so that it recognizes the same item across several suppliers
• units and variants cleanly separated from each other
• up-to-date price and availability from the ERP
• texts and media in every language you sell in

None of these points can be solved in the feed alone, because each one originates in one of the systems behind it.

Most of what an agent needs to read is defined by product data management in the PIM system. This is where it is specified which attributes a product group has, which data type and unit each attribute carries and which values are permitted.

This data model determines whether an agent can tell units and variants apart. If “25 mm” is entered in a text field, it remains unclear whether it refers to length, diameter or nominal diameter. As an attribute with a numeric value and its own unit, the same information is unambiguous. Variants work in a similar way, because an item “available in three sizes” can only be ordered once each size has its own item number.

Translations are also handled in the PIM with an approval workflow, so that every language reflects the same status. If a language is missing, the assistant may translate on its own, and its technical terms will then no longer match the data sheet.

For the same reason, values that an AI tool adds in the feed belong back in this place. Product data management transfers enriched values into the PIM system, checks them there and approves them before the feed and shop pass them on to the assistants. If they remain only in the feed, they sit outside product data management, and the assistant reads “stainless steel” as the material while the field in the PIM is empty. How the same data supports search, recommendations and categorization in your own online shop is described in the article AI Needs Clean Product Data.

Prices, stock levels, delivery times and terms originate in the ERP and change there constantly. If an agent reads an outdated stock level from the feed, it recommends an item that is not available, and purchasing only notices when the order confirmation arrives. Feed and shop therefore read these values in real time from the ERP instead of maintaining their own data.

In B2B, prices are also often customer-specific. In MS Dynamics 365 Business Central, for example, price lists, discount groups and agreements determine what a particular customer pays, while a feed shows the list price at best. If an agent is to order on behalf of a customer in the future, it therefore needs access to that customer’s account or to an interface that provides price and availability for exactly that customer. The order would then arrive in the ERP as an order from that customer, just like an order from the online shop. There is currently no established procedure for agents to authenticate for this, but one is sure to come.

This is most likely to emerge first for reordering consumables, such as seals or cleaning agents. An agent would check stock and the framework agreement, create the shopping cart and obtain approval before placing the order.

There are also organizational decisions to be made. It would need to be clarified which agents may order on behalf of which customers, up to which order value, and who approves when the agent suggests a substitute item. Many B2B shops already know such rules from customer accounts in which several employees order with different permissions. So the foundation has often already been laid.

A test with a handful of typical items shows how far your own data already goes, including one item with variants and one that sales is often asked about. An AI assistant such as ChatGPT, Gemini or Copilot is given the questions customers ask about each item, for example about material, dimensions or delivery time. In doing so, it reads the same data that an agent will later need to place an order. Where an answer is wrong or missing, four criteria lead to the cause:

• The most important technical information is maintained as an attribute with a unit and does not only appear in the description text.
• Each size or version has its own item number.
• Price and delivery time in the feed match the status in the ERP.
• The item is complete in every sales language.

For each gap, note which system the correct value should be stored in. This shows which gaps product data management closes in the PIM and which require a connection to the ERP.

None of the requirements in this article is new. A good dealer portal already needs exactly the same today, namely attributes for filters, customer-specific prices after login, up-to-date availability and content in every sales language.