A combination of RAG and FineTuning will be much more useful IMHO.
Think of LLMs as generic models that can answer anything but with lower accuracy.
You fine-tune them to learn specifics of a particular domain. This way LLMs can provide more factual answers to domain specific questions.
Finally, you can add RAG on top of fine tuned models to get answers in context of your organisation or specific documents.
How all this will pan out remains to be seen, but surely there are interesting applications to come out of these technologies.
