Untangling Product Classification with AI

We’re working with an international retailer that operates its own tills, infrastructure, and product coding, but sells within larger third-party environments across multiple countries. In these settings, commissions are calculated not only by sales value but also by product type, according to external classification hierarchies set by the host operator.

These hierarchies vary significantly from site to site and are often written in local languages. In some cases, different rules even apply within the same site depending on the context of the sale. While our client uses its own internal product categories, these need to be mapped to each external classification system. This process becomes increasingly fragile at SKU level, leading to mismatches, disputed commissions, and a growing burden of manual intervention.

To address this, we’re helping our client build the capability to apply external classifications directly at SKU level and to get it right during product setup rather than after the fact. This involves asking product buyers to classify items using external logic that is often unclear, inconsistently documented, and written in another language.

Processing batches of 100 products at a time, we trained an AI model to predict the appropriate classification based on internal data points like product descriptions and categories. After several rounds of training and refinement, accuracy rose from 50 percent to over 90 percent.

It worked well. Incorrect classifications were typically marked Low, while accurate ones showed up as High. This gives buyers clear guidance on where human review is needed and where automation can speed things up.

The result is a practical AI application that enhances product setup, improves classification accuracy, and reduces downstream friction. It’s a strong example of how AI in retail can be deployed with real impact.