What it means in practice
Buy a suit off the rack and it usually fits reasonably well. If you want it to fit perfectly, you take it to a tailor who adjusts it to you.
That is exactly what fine-tuning is. You take a finished, capable LLM and show it thousands of examples from your company. Five thousand of your old customer emails, for instance, so it learns the tone your company uses, the jargon you use, and how you format your replies.
Why companies need to know this
The most common misunderstanding we meet among owners is: “We want the AI to know our internal policies, so we have to train it on them.” That is a mistake.
- For facts, use RAG. If you want the AI to know your prices, your returns terms or what is in the latest contract, fine-tuning is the wrong and expensive route. Models learn facts poorly and unlearn them worse when those facts change. Facts and document search are what RAG is for.
- For style and tone, use fine-tuning. Fine-tuning is ideal when you want to change how the AI speaks. If you are a law firm and want contracts written in your specific legal style, fine-tuning on your old contracts is the right route.
- Cost and data. Fine-tuning is not cheap. It requires hundreds or thousands of good examples (a dataset). Pour rubbish into the model and it learns to produce rubbish.
An example from practice
An e-commerce company wanted AI to write product descriptions for their shop in their specific, slightly cheeky and funny tone.
They tried long instructions first, but the AI kept sounding robotic. So they took a thousand of their best existing product descriptions and fine-tuned a smaller, cheaper model on them. The result: the model learned their exact company voice and now produces descriptions the copywriters barely have to touch.
In our company AI workshops we help companies decide whether a good prompt solves their problem, whether they need RAG for document search, or whether investing in fine-tuning genuinely pays.