What is fine-tuning? (and why you probably don't need it)

Fine-tuning is the option businesses ask about first and need least. It sounds like the serious answer — a model trained on your data, tailored to your business. In practice it's an expensive solution to a problem that usually has a cheaper one.

What it actually does

A general model has been trained on an enormous amount of text. Fine-tuning continues that training with a much smaller set of examples you supply — say, five hundred pairs of "here's the input, here's the output we want".

The model adjusts toward those examples. Give it hundreds of your support replies and it starts writing in that register without being asked. Give it hundreds of correctly categorised records and it internalises your categories.

What changes is disposition, not knowledge. Think of it as training a new employee in your house style rather than telling them about this month's projects.

The distinction that matters

You want…What you actually need
It to answer using our documentsAccess to those documents
It to know our current pricesAccess to the price list
It to always write in our toneA system prompt — then maybe fine-tuning
It to output the same rigid structure every time, at volumeFine-tuning is a fair candidate

Fine-tuning is also the worst possible way to give a model current information. Anything you train in is frozen at training time — change your prices and you'd have to train again.

An Australian small-business example

A Sydney insurance brokerage wants AI drafting client emails that sound like them. They're quoted a fine-tuning project: gather examples, prepare data, train, evaluate. Weeks of work.

Instead they write a system prompt: the tone they want, the phrases they never use, the compliance line that goes at the bottom, and three example emails pasted in. Twenty minutes.

It gets them roughly ninety per cent of the outcome. And when their compliance wording changes, it's a one-line edit rather than another training round.

Fine-tuning, RAG and prompting, side by side

That last point catches people. Fine-tune today, and when a materially better model ships in six months you're choosing between your tuning and the upgrade.

The honest recommendation

For nearly every Australian small and medium business, fine-tuning is a distraction. The gains people hope for come from three cheaper things: writing better instructions, making those instructions standing rather than repeated, and letting the model see the real documents.

Exhaust those first. If you've done all three properly and still have a narrow, high-volume task the model won't do consistently, fine-tuning becomes a reasonable conversation — with a clear view of what it will cost to maintain.

Frequently asked questions

Will fine-tuning teach the model about our business?
Not reliably, and this is the most common misunderstanding. Fine-tuning shapes behaviour — tone, format, structure. For facts about your business, you want the model reading your actual documents, which is a retrieval problem, not a training one.
How much does it cost?
The training run itself is often modest. The real cost is preparing hundreds or thousands of high-quality examples, and then redoing that work each time you want to change something or move to a newer model. That maintenance burden is what catches people out.
When is fine-tuning actually the right answer?
When you need a consistent, narrow output format at high volume and prompting has genuinely failed — classifying thousands of records into your own categories, say. It's a specialist tool for a specialist job, not a general upgrade.
What should we try first?
In order: a better prompt, then a system prompt that sets standing rules, then giving the model access to your real documents. That sequence resolves the overwhelming majority of cases, and each step costs a fraction of fine-tuning.
Does fine-tuning make the model smarter?
No. A fine-tuned smaller model can beat a general larger one at one narrow task, but it doesn't gain general capability — and it often gets worse at everything outside the task you tuned it for.

Put this to work

Ad On Group runs AI training and enablement for Australian teams through Ad On AI — a three-month, self-paced program that takes non-technical staff from their first prompts to working AI agents.

Talk to us →

Keep reading

← All resources