"Fine-tune a model" is the request we get the most — and the one we turn down the most. Here's why, and the three steps to climb first.
By Nacim Moudjeb6 min5
The request we get the most — and turn down the most
"We'd like to fine-tune a model on our data." We hear it every week. And most of the time, we say no. Not for lack of skill — we do this. Because nine times out of ten, it's a €30,000 answer to a problem that a few hundred euros would solve.
"Fine-tuning" has become a reflex. It sounds serious, tailor-made, premium. The trouble is people credit it with a power it doesn't have. Before signing off on a project like that, you need to know what it does — and above all what it doesn't.
What fine-tuning actually does
Fine-tuning means taking an already-trained model and showing it thousands of examples to adjust its behavior. The result: the model answers differently. It picks up a tone, a format, a manner. It's good at one thing — form.
What fine-tuning does not do is learn your data. That's the number-one misunderstanding. Fine-tuning a model on your 4,000 product sheets doesn't teach it your products — it teaches it to imitate the style of your sheets. Ask it the price of a specific item and it will invent a plausible answer. With full confidence. Exactly what you don't want.
To make a model actually know your data, there's another method. Simpler, cheaper, more reliable.
The three steps to climb before fine-tuning
The rule, shared by pretty much everyone who does this seriously (IBM and InterSystems put it the same way): try prompting, then RAG, and only fine-tune as a last resort. Three steps. In the vast majority of cases, you never go higher.
1. Prompting. Give the model clear instructions, examples, the context of the task. It sounds basic. It's wildly underrated. A well-built prompt turns a generic model into an assistant that answers like your best employee — without a single line of training. Plenty of "we need to fine-tune" projects get solved right here, in two days.
2. RAG: plug in your data. RAG connects the model to your documents, your CRM, your pricing, live. When asked a question, it fetches the right information from you, then answers with it. It quotes your real numbers instead of inventing them. That is "an AI that knows your business" — not fine-tuning. And the day a price changes, you update a document, not an entire model. See our approach to RAG and knowledge bases.
3. The agent: give it hands. A model on its own only talks. The agent is everything you build around it so it can act: read your tools, run actions, check its own work, try again. The difference between an AI that explains how to make a quote and one that makes the quote. Almost all the value we install for clients comes from here, not from a retrained model. That's the whole point of our AI agents.
So who is fine-tuning for?
For real but narrow cases. When you need a very specific tone or format, repeated millions of times, that prompting can't hold reliably enough. When volume is so high that shortening each request saves real money. When a niche task — classifying documents in a very specific trade, say — genuinely benefits from being baked into the model.
These cases exist. We handle them; it's even one of our services. But they come after you've proven value with the three steps above — not instead of them. Fine-tuning is optimization, not a starting point.
The real cost of fine-tuning too early
Fine-tuning isn't "train once." It's building a clean dataset — the bulk of the work, and the most thankless — then training, evaluating, hosting the model, and starting over every time your data or your needs shift. You're no longer maintaining a prompt: you're maintaining a model.
Meanwhile, the competitor who plugged a good RAG into a recent model has already shipped — and gets every new version of the base model for free. You're frozen on the version you trained. That's the trap: paying more to move slower.
Where to start
We take your case and climb the steps in order. Prompting first. Your data via RAG if needed. An agent if the task calls for acting, not just answering. And fine-tuning only if, at the end, it's the one step still missing. It's less of a sales pitch than "we'll train your AI." It's mostly what works, and what costs you the least.
If someone offers to fine-tune before trying the rest, be wary: they're either selling what they know how to do, or billing by the kilo.
Want us to look at your case without selling you the step above? Book a free audit.