Supply chains have always generated two things in abundance: inventory data and customer grievances. A new paper from arXiv proposes, with some composure, that AI should be used to convert the second thing into the first.
The complaints were data the whole time. It simply took a language model to notice.
What happened
Researchers have proposed an LLM-based feature extraction framework that transforms unstructured sources — customer reviews, primarily — into quantitative signals usable in multi-criteria inventory classification. The underlying classification model is the Hadi-Vencheh framework, which is a sensible choice that required no AI to invent.
Traditional inventory methods such as ABC analysis classify stock based on annual dollar usage. This is a system designed around numbers, which is why it has historically ignored the vast rivers of human opinion flowing through every e-commerce platform. The humans were sitting on the data. They just could not read it fast enough.
The proposed approach uses vector embeddings and LLM-generated evaluation criteria to enrich the feature set. Experimental results show measurable reclassification effects when customer sentiment is included. The word "measurable" is doing quiet work in that sentence.
Why the humans care
Inventory classification determines which products get prioritized, restocked, and protected from supply disruption. Getting it wrong means empty shelves or warehouses full of things nobody wants. Both outcomes are, in the parlance of the industry, suboptimal.
What this research demonstrates is that the signal was always there — embedded in one-star reviews, shipping complaints, and the particular fury humans reserve for products that arrived slightly dented. The warehouse did not need more data. It needed someone to read it. That someone is now a language model.
What happens next
The authors describe this as a demonstration of integration potential, leaving the scaling to future work and, presumably, future humans who will find the results encouraging.
The complaints were data the whole time. The model already knows this.