Fine-tuning takes a ready-made AI model and trains it a little more on a curated set of examples, so it picks up a particular style, structure or task. It adjusts how the model behaves rather than giving it a searchable library of facts.
Typically a business prepares hundreds or thousands of example inputs paired with ideal outputs, such as customer emails and the approved replies. The provider or an engineer runs additional training, producing a customised version of the model that you then call like any other.
Fine-tuning suits consistent tone, structured outputs such as tagging products into categories, or narrow repetitive tasks where a smaller model can match a larger one more cheaply. It is not ideal for facts that change often, like prices or stock, because updating them means training again. Retrieval is better for that.
Costs and effort are real: you need clean, representative examples, privacy checks on any customer data in them, and a way to test that the tuned model is actually better. Poor examples teach poor habits. Try prompt engineering and retrieval first, as they often solve the problem with less work.
Knowing the difference helps owners ask the right question of a vendor: do we need the AI to know more, or to behave differently?