You supply pairs of inputs and the outputs you wanted, and training adjusts the model so it becomes more likely to produce answers like your examples and less likely to produce the ones you'd have to correct.
It resembles a well-read new hire spending their first month in your office. They could already write; the month teaches them your formats and standards, though not every fact about the business. Providers document the process and its limits, for example in the OpenAI fine-tuning guide.
What is fine-tuning good for?
Fine-tuning is good at behaviour: tone, structure, classification, a fixed output format, or a narrow task the model keeps doing slightly wrong in the same way. It's poor at knowledge, because facts change and the training is frozen at the moment it was done. If the model doesn't know your current price list, fine-tuning won't help much; giving it the price list is cheaper and stays accurate.
An example
A bookkeeper processes several hundred receipts a month and wants each turned into the same line: date, supplier, net amount, VAT, and one of twelve expense categories from her client's chart of accounts. With instructions in a prompt, the model gets most of them right, but it keeps filing software subscriptions under "office costs" and dropping the VAT field when the VAT is zero.
Fine-tuning on a few hundred corrected examples fixes both habits, because the pattern is stable and the output shape never changes. It would do nothing for a question like "what's the current VAT registration threshold?", which changes by government decision and belongs in a supplied document or a quick check on gov.uk.
What does a fine-tuned model cost to keep?
A tuned model is tied to one base model and one set of examples. When the provider retires that base model, or your task changes, you prepare the examples again and retrain. That ongoing cost rarely appears in the pitch, and it's why most teams try better prompts and supplied documents first.
Fine-tuning or RAG?
If the model writes in the wrong shape or style, tuning can help. If it lacks a fact, retrieval-augmented generation supplies that fact at the moment of answering. When you're unsure, start with a good prompt and the right documents, because both are cheaper and quicker to change.
Common questions
Is fine-tuning expensive?
It costs more than prompting, both in money and in the time it takes to prepare clean examples, and it has to be redone as the task changes. For most small businesses, better prompts and better context give more improvement for less effort.
Can fine-tuning teach a model new facts?
Partly, but unreliably, and those facts start going stale the day training ends. Facts are better supplied at the moment of use, where they can be updated and cited. Tuning also doesn't stop hallucinations.
Does fine-tuning make a model safer?
Not by itself. It can make behaviour more consistent, but safety comes from limiting what a system can reach and reviewing what it does, and training alone provides neither.
Does a personal agent need fine-tuning?
Rarely. A personal agent picks up your preferences through memory and saved instructions, which you can change in a sentence. A fine-tuned model has to be retrained to change its habits.