Fine-tuning takes a model that already understands language and trains it further on a focused set of your own examples. You are not building intelligence from zero; you are nudging a capable model toward a narrower job, such as always replying in your brand voice or emitting perfectly structured JSON. Providers expose this through an API where you upload examples and receive a private version of the model.
The practical trigger for fine-tuning is repetition. If you keep pasting the same lengthy instructions into every prompt, baking that behavior into the model can shorten prompts, cut cost per call, and improve consistency. It shines for style, format, and classification tasks that are hard to describe but easy to demonstrate.
It is the wrong tool for teaching facts. Fine-tuning changes how a model behaves, not what it knows, and models do not reliably memorize details you show them this way. When you need the model to reference specific, changing information, retrieval-augmented generation is the better and cheaper answer.
Budget for the hidden work: collecting clean examples, holding some back to test against, and retraining whenever the base model updates. A few hundred high-quality examples usually beat thousands of sloppy ones. Start with prompt engineering, and only reach for fine-tuning once you have proven the ceiling of what prompts alone can do. That sequencing keeps your effort where it pays off.