AI auto-classification: assigning products to ETIM, ECLASS and GPC

What does AI auto-classification do in a PIM?

AI auto-classification assigns each product to a class in a taxonomy, using its name, description and attributes. The taxonomy can be internal, a trade standard such as ETIM or ECLASS, or a channel category tree. The model suggests; a product data steward decides.

PIMvendors rates auto-classification as production-ready with review in the AI in PIM maturity table. The target list is finite, so the model chooses among known options. The steward can check a suggestion against the class definition in seconds.

Classification matters more than its size suggests. In most data models the class determines which attributes a product needs and which values are allowed. A wrong class assigns the wrong attribute template. The product then misses required data, appears under the wrong filter and fails validation at data pools and retailers.

Three situations create the strongest case for AI classification:

• Onboarding large supplier assortments that arrive without classification.
• Adopting a trade standard such as ETIM or ECLASS for an existing catalogue.
• Mapping one internal taxonomy to several channel category trees.

In each case the work is repetitive, the options are known and a specialist can review faster than classify from scratch.

Which classification systems can AI target?

AI can target any classification system with defined classes, but each system has its own structure and users. Choose the systems your channels and customers require before testing.

The standards differ in depth. GS1 describes GPC as building down from segments to bricks, with brick attributes for product details. One GS1 publication reported 1,636 bricks, 3,144 attributes and 25,081 values. ETIM classes carry features with defined values and units; the ETIM page explains the model.

Channel trees are a separate task. OpenAI's product feed, for example, accepts a category path with levels separated by the > character. Map your internal classes to each channel tree once and let the PIM apply the mapping on export.

How accurate is AI classification and how do you measure it?

Measure AI classification with top-1 and top-3 accuracy per class, because overall accuracy hides weak classes. A model can score well on a catalogue full of easy classes and fail on the technical ones that matter.

Neighbour confusion is the error to watch. Classes with similar names, such as different types of cable glands or connectors, attract the most mistakes. These errors also pass review most easily, because the suggestion looks plausible.

An abstention option helps. A model that says it cannot classify an item with confidence saves the reviewer from a confident wrong answer. Ask vendors whether their classification can abstain and route the item to manual work.

How should the review step for AI classification work?

The model proposes three ranked classes with confidence scores, and a steward selects one. That design keeps the speed of AI and the judgement of a specialist.

1. The model returns the top three classes with confidence scores.
2. Items above a calibrated threshold go to bulk review, grouped by suggested class.
3. Items below the threshold go to individual review.
4. The steward selects a class, or marks the item for supplier clarification.
5. The PIM logs the suggestion, the selection and the reviewer.

Grouping by suggested class speeds review. A steward scanning 200 items suggested for one class spots outliers faster than one reviewing 200 mixed items.

Corrections should feed back. When stewards repeatedly override one suggestion, that pattern reveals an unclear class definition or a model weakness. Ask vendors whether corrections improve later suggestions and how quickly.

Classification also feeds AI attribute extraction. Once the class is set, extraction knows which attributes to look for. Run classification first, then extraction, and test both in the same round.

What happens when a classification standard releases a new version?

A new standard version triggers a reclassification project, and AI can reduce it to a class-mapping review. ETIM 10.0, for example, was released in December 2024, and companies on earlier versions face class and feature changes.

A manual reclassification touches every product. An AI-assisted approach works at class level:

1. The model proposes a target class for each old class, using the release's change information and product examples.
2. A steward approves the mapping once per old class.
3. The PIM applies the mapping in bulk to all products in that class.
4. The model flags products that do not fit the mapped class for individual review.
5. New or changed features become attribute gaps in the completeness report.

Review effort then scales with the number of classes in use. For a catalogue with tens of thousands of products in a few hundred classes, that difference is large.

Version handling belongs in your requirements. Ask whether the PIM can hold two versions of a standard during transition. Ask too whether it can export the older version to customers and track which products are migrated. The data model design page covers versioned classifications.

What should you ask vendors about AI classification?

Ask vendors to classify a sample of your products into your target systems and report top-1 and top-3 accuracy per class. Demo classifications on clean sample data predict little.

Requirements to include:

• Classification into internal taxonomies and the trade standards you use, per version.
• Top three suggestions with confidence scores and an abstention option.
• Bulk review grouped by suggested class.
• Logging of suggestion, selection and reviewer per product.
• Support for class mapping during standard version upgrades.
• Mapping from internal classes to channel category trees.
• Pricing per classified item and per reclassification run.

Prepare the test set carefully. Include products from every major class, a share of hard cases from neighbouring classes and some items that should not fit any class. Hold back the answer key. The AI Claims Lab protocol includes classification as one of four tasks on 200 buyer-provided products and provides the scoring sheet.