Why use an LLM as opposed to a more narrow purpose built model? LLMs are not beating smaller, purpose built models on tasks like POS tagging, NER, sentiment analysis, etc. And the inference costs scale quite poorly (unless you are self hosting llama or something).