Prompt, RAG, fine-tuning: three words that most managers confuse, and this confusion is costly. What your SME really needs is rarely fine-tuning. I explain to you
The scene comes up on almost every first date. A manager told me: “We need an AI trained on our data. » He read somewhere that this was the serious stage, the one that separates DIYers from real projects.
In 9 out of 10 cases, I tell him to wait. Because what he describes often costs 20 times more than the solution he actually needs, and comes out 4 months later.
Three words come up whenever we talk about adapting AI to a company: prompt, RAG, fine-tuning. The majority of decision-makers confuse them, and this confusion is paid for in euros and lost months. Here’s how to untangle them in 5 minutes, and how to know which one concerns you.
The prompt: 80% of your needs, for 0 € model
The prompt is how you speak to the model. The instructions you give him, plus the context you put in front of him when asking the question.
It seems basic. However, this is what resolves the overwhelming majority of cases in SMEs.
A concrete example. An accounting firm wanted to automatically sort its client emails by type of request. No project at 30,000�� �1. We wrote a precise prompt, pasted 5 examples of emails already filed, plugged it into their inbox. In production in one week, for the price of the API subscription.
When the prompt is enough: write quotes, summarize reports, respond to standard emails, reformulate, extract information from a document. If your need is “explain to the model what you want and give him one or two examples”, you don’t need anything else.
The RAG: when AI must read your documents
RAG means “recovery augmented generation”. The name is scary, the idea is simple.
Imagine an assistant who you don’t ask to memorize your 4,000 documents. You stick a librarian next to him. For each question, the librarian will search for the 3 useful passages in your files and place them on the desktop. The AI responds from that.
You need it when the AI must rely on your subject: a support chatbot connected to your knowledge base, a search in hundreds of contracts, an assistant who responds to your internal procedures.
The big advantage: when you add a document, it is taken into account immediately. No reprogramming. You have control over what the AI is allowed to quote, and you know where each answer comes from.
Allow a few weeks of implementation depending on the volume. This is the level above the prompt, and it covers almost all of the remaining cases.
Fine-tuning: what everyone asks for and almost no one uses
Fine-tuning means re-training the model on your own examples so that it permanently adopts a precise format or tone.
Here, we touch on the real “training on your data” that managers ask for. And that’s where I brake, almost every time.
For it to be worth it, three conditions must be met at the same time: a very narrow and repetitive task, a large volume, and thousands of clean and well-labeled examples on hand. Most SMEs don’t have any of the three.
The trap that we rarely see coming: the day your process changes, you have to re-train everything. And every few months a new base model comes out that’s better than your specialty version from six months ago. Your investment ages quickly.
And the most expensive thing about it is data preparation: a long and manual job that no one likes to do.
The bad reflex that is expensive
The pattern I see most often: a company jumps straight to fine-tuning because it sounds serious, and skips the prompt and the RAG.
Current result: €40,000 and 4 months to obtain something that an assembly at €200 per month would have delivered in 2 weeks. With, as a bonus, a fixed model that will have to be retrained at the first evolution.
The correct order is the reverse. We start with the prompt. If it blocks, we add RAG. Fine-tuning comes as a last resort, once you have proof that nothing else is enough.
How to choose, in 3 questions
Ask yourself these questions in order.
- Does the pattern just need good instructions and an example or two? So it’s prompt. Stop there.
- Should he draw on your up-to-date documents and cite his sources? So it’s RAG.
- Does it have to reproduce a very precise format or tone, thousands of times a day, with examples already labeled on hand? Only there, the fine-tuning is discussed. And again: test the RAG first.
If you’re stuck on the first question, this is already excellent news for your budget.
Where to put your money
The inconvenient truth about great slides: the model is rarely the problem.
Your useful budget goes elsewhere. In the storage of your data (a clear customer file is better than a spiked model). In well-written and tested prompts. And in the piping, the connection of AI to your existing tools, your CRM, your mailbox, your files.
A starting SME needs a good prompt, sometimes a little RAG, and someone who connects all of that to their daily life.
The day you really need fine-tuning, you’ll know it. Your data will tell you, and so will your volume. In the meantime, keep the €40,000.
