Why it matters
Mostly, it matters because you probably don't need it, and knowing that saves money. Vendors love to sell "your own custom model"; in practice, good prompting plus RAG covers most business needs at a fraction of the cost and complexity.
How it works
You assemble hundreds or thousands of example input-output pairs showing exactly the behaviour you want, and train the base model further on them. The result is consistency at volume for narrow tasks (always this structure, always this tone) at the price of dataset building and maintenance.
What to do about it
Exhaust prompting and RAG first. Fine-tune only when a narrow, high-volume task still misses the bar and you have a pile of good examples to teach from.
