What is fine-tuning in AI?
Fine-tuning continues training an existing model on your own examples so it learns a style, format or task, instead of being re-instructed through the prompt on every call.
In more detail
Fine-tuning changes the model's weights. It is good at teaching consistent behavior: a fixed output format, a classification scheme, the tone of your brand. It is poor at teaching facts, because knowledge baked into weights is hard to update and hard to cite. For facts that change, use retrieval instead (see RAG). A typical project needs a few hundred to a few thousand labeled examples, a held-out test set and a plan to retrain when the task shifts. Often a better prompt, or a few examples inside it, gets most of the way, so we try that first and fine-tune only when volume or cost justifies it.
Example
An insurance back office fine-tunes a small model to sort claim documents into 14 types. The team compares it with a prompt-only baseline on a held-out set and adopts it only if it wins on accuracy and cost.