Fine-tuning vs. prompting vs. RAG: which to use

Throughout 2023, as language models went from curiosity to daily use, a technical debate came to a head that’s still alive today: how do you get an AI to answer with what’s yours —your products, your policies, your history— and not with generalities pulled from the internet? The conversation organized itself around three paths: prompting, RAG, and fine-tuning. They sound like jargon, but at their core they answer a very business-minded question: how to teach the machine your company’s context.

The good news is that you don’t have to choose blindly or overpay. Each path solves a different problem, has a different cost, and fits a different moment in your operation. In this guide we explain all three in plain language and, above all, which one makes sense for a Mexican small business that wants measurable results without putting its data at risk.

Prompting: give it clear instructions

Prompting is simply telling the AI clearly what you want, with examples and instructions within the same conversation. You don’t modify the model: you give it context in the moment. It’s the fastest, the cheapest, and, for many tasks, all you need.

• When it works: drafting emails, summarizing documents, classifying messages, generating drafts. Tasks where the model’s general knowledge is already enough.
• Advantage: you start today, with no infrastructure or long projects.
• Limit: the model doesn’t “remember” your catalog or your policies between sessions; you have to give it the context each time.

For most small businesses, this is where it all begins. Before investing in something more complex, it’s worth squeezing everything a well-designed prompt already gives you.

RAG: let it consult your own information

RAG (retrieval-augmented generation) connects AI to your documents: manuals, catalogs, contracts, ticket history. When someone asks, the system first searches your information and then answers based on it. It’s like giving AI access to your filing cabinet, without it having to memorize it.

• When it works: an assistant that answers about your products, an internal policy search, customer service based on your real documentation.
• Advantage: answers rest on your information and update when you update your documents. It keeps your data, processes, and history as the source of truth.
• Consideration: it requires organizing and maintaining your information, but it’s an investment that pays off in any process.

For almost any small business that wants an AI “that knows the house,” RAG is usually the sweet spot between effort and result.

Fine-tuning: adjust the model to your style

Fine-tuning retrains the model with your own examples so it adopts a very specific tone, format, or behavior. It’s the most powerful path and also the most costly: it’s worth it when you already have volume, quality examples, and a need that prompting and RAG don’t cover.

The most common mistake isn’t choosing the wrong technique, but starting with the most expensive one before exhausting the simplest.

It’s rarely a small business’s first step. It makes sense when the business has already validated the use case and needs to fine-tune the last mile of quality or consistency.

How to choose your path

Our recommendation at Normandia Web is to move forward in layers, measuring at each one:

• Start small: choose a single process with real pain —service, quotes, internal support— and set a clear goal.
• Test with prompting first: it often solves 80% without investing in infrastructure.
• Move up to RAG when you need your data: if the AI must know your catalog, policies, or history, that’s the natural jump.
• Reserve fine-tuning for last: only when you’ve already validated the case and need to fine-tune tone or consistency at scale.
• Take care of your data: define from the start what information goes in, where it lives, and who sees it.

The decision isn’t “which technique is better,” but which one solves your problem at the lowest cost today, while leaving the door open to grow. With custom software you can combine prompting, RAG, and fine-tuning as your operation matures, without getting locked into a single tool. If you’d like, at Normandia Web we’ll sit down with you to map your case and choose where to start.