I did something similar 10 years ago, but instead of llms I used word2vec to calculate a embeddings of product descriptions and map those to existing categories. The LLM approach is very likely better, but I'm curious what the cost difference is.