Why it is harder than it looks
Existing classifications are unreliable: catalogues inherit years of ad-hoc filing, and the right category is often in a different branch from the original one. A smoke detector lost in the wrong category remains invisible until someone looks where it should be. Choosing among hundreds of categories based only on their labels is another trap: two similarly named categories may contain very different products.
A second opinion only has value if it is blind
The classic mistake when checking an existing classification is to show the current category to the AI “to help it”. The measurable result is anchoring bias: the agent confirms the existing choice instead of reassessing it, the “confirmed” rate rises and real errors slip through. The right method is blind categorisation. The agent receives the product without its current category or original branch, identifies it through web research and manufacturer data, then classifies it from scratch. This is the condition for a genuinely independent second opinion rather than another confirmation of the status quo.
Check with examples, not labels
A category label does not reveal what the category contains. Before validating a choice, ask a concrete question: “is this product the same type of object as the products already filed in this category?”. The agent examines ten or twenty real products from the candidate category and compares them with those in the current category. Two arbitration rules do the rest: always prefer the most specific category — “filter elements for air treatment” rather than “air treatment” — and, when both are equally specific, do not change. Reclassification needs a clear reason.
Confidence scores — here, and only here
Categorisation is the one place where a numerical confidence score makes sense: the choice is closed — one category among N — and therefore measurable. A simple scale is enough, and every suggestion includes a plain-language reason:
| Level | Meaning | Action |
|---|---|---|
| Near-certain (≥ 0.80) | Web research agrees with products already classified in the category. | Can be validated in bulk after sampling. |
| Probable (0.50–0.79) | Only one signal, or ambiguity between two similar categories. | Quick review using the reason provided. |
| Weak hypothesis (< 0.50) | Product is difficult to classify or data is insufficient. | Mandatory human review — never forced. |
Conversely, a “confidence score” attached to each enriched value is an artifice: an extracted value is proven by its source, not a percentage.
Classify groups, not individual products
A category is a property of the model, not the variant: all thirty-six variants of a jacket belong to the same category. Once the catalogue has been grouped into parent/variant structures, the group is categorised by reading all its variants. Seeing the whole group reveals the model far more clearly than one isolated product, and the category propagates at once. One safeguard remains: a member that clearly is not a variant of the model is flagged as an outlier rather than inheriting the group's category. Categorisation also checks the grouping as it proceeds.
Customer taxonomy and market taxonomies
The same method classifies products within the customer's own taxonomy — its category tree and departments — or within market classifications: GS1, or ETIM for technical distribution. Once the category has been validated, it carries consequences: the expected category attributes become the enrichment schema. Classification is not an end in itself; it is the contract for everything that follows.
Related topics: parent/variant · product data enrichment · product data quality