Can customer reviews change which inventory deserves attention?

A warehouse cannot give every product the same level of attention. One familiar way to set priorities, called ABC analysis, ranks items by their annual dollar usage and divides them into three classes. That makes financial importance visible, but it does not directly account for the delivery experience described in customer reviews.

In a preprint, Humza Akhtar and Rami Pinto propose a way to bring such text into a broader inventory classification. They compare the resulting classes with those from two approaches that use only structured numerical information. Their central result is a change in classification, not proof that a business would perform better by following the new rankings.

From words to a score

Multi-criteria inventory classification extends ABC analysis by taking several attributes into account. The researchers' conventional comparison uses annual dollar usage, average unit cost and lead time. Their augmented version adds a fourth attribute: perceived delivery experience, derived from customer reviews.

The proposed process first defines a criterion for the information of interest. A language model helps translate a business objective into a scoring rule, while domain experts supply the objective and validate the criterion. In this case, more negative delivery experiences receive higher scores, reflecting a greater potential need for attention; positive experiences receive lower scores.

The framework represents text as embeddings: numerical representations intended to capture aspects of meaning. It uses these representations to retrieve relevant evidence for a product and criterion, then uses a language-model-assisted evaluation to score the retrieved context. The resulting score joins the structured attributes in a product's feature set. The authors describe implementing the framework with MongoDB and its vector search.

The classification model then normalizes the criteria and calculates a composite score. Rather than requiring experts to specify exact weights in advance, the Hadi-Vencheh model constrains weights using an expert-defined ordering of the criteria. Products are ranked and assigned to A, B or C priority classes. The point of the exercise is to see how the extra review-derived criterion changes those assignments.

How much did the classes move?

The evaluation uses the public Brazilian E-Commerce dataset from Olist, which the paper describes as containing more than 100,000 anonymized orders alongside product, delivery and review information. Orders are not the same thing as the inventory items being classified, so that figure should not be read as the number of classified products.

When the researchers compared conventional multi-criteria classification with the version incorporating delivery experience, 69.43% of items remained in the same class. Another 14.51% moved to a higher-priority class and 16.06% to a lower-priority class. Those are changes relative to the conventional multi-criteria result, not evidence that each individual move was correct.

The paper also compares the two multi-criteria approaches with traditional ABC classification. Adding structured criteria alone left 63.12% of items in the same class as ABC; comparing ABC directly with the review-augmented approach left 55.19% unchanged. That last comparison combines the effects of the structured criteria and the review-derived signal, so it cannot isolate the influence of reviews.

What the experiment cannot establish

A more negative review-derived score is treated as a reason for closer attention by design. A higher priority could identify an item for managers to investigate, but a changed class is not itself an observed improvement in delivery, service or stock availability. The authors explicitly say their results show reclassification rather than validated gains in operational performance.

The paper calls for testing against outcomes such as service levels, stockouts and holding costs. It also identifies robustness checks and bias detection for language-model-generated scores as requirements for practical use. Those checks matter because the new ranking depends not only on the reviews available, but also on how evidence is retrieved, how the criterion is defined and how the model scores it.

The study illustrates a way to put customer feedback into an inventory decision that would otherwise use numerical attributes alone. Whether those different decisions are better remains an open empirical question.