# What is fine-tuning an AI model, and when is it worth it? > What fine-tuning an AI model is, how it works, what it’s for and what it isn’t, how it differs from RAG and from the prompt, and what you need to do it. Joaquín Tenés · 2026-11-14 · https://delegic.com/en/glossary/fine-tuning “Can we train the AI on our data?” is one of the first questions a company asks when it thinks about an agent. Fine-tuning is, literally, that: training a model on your examples. But it’s almost never what’s needed for an agent to know your rates or who’s on call. Here’s what it is, how it works, what it’s for, what it isn’t for and when it’s worth it. ## In short - Fine-tuning is continuing to train an existing model with examples of a specific task. - It changes the model inside, its weights; the prompt and RAG only change what it reads with each answer. - It’s for teaching a way of answering: a tone, a format, a task that repeats a lot. - It isn’t for making it know facts that change, like rates, on-call rotas or a customer’s contract; Microsoft says so explicitly. - It has training and maintenance costs, and its benefits aren’t guaranteed: they have to be measured. ## What is fine-tuning? A [language model](/en/glossary/llm) is first trained on an enormous amount of general text. Fine-tuning is a second, much smaller round of training that starts from that ready-made model and adjusts it with examples of a specific task. [IBM](https://www.ibm.com/think/topics/fine-tuning) defines it as the process of adapting a pre-trained model for specific tasks or use cases, and the point is not starting from scratch: everything the model already knows is kept, and it’s only taught what’s missing. The difference from the other ways of adapting a model is where what you teach it goes. [Microsoft’s documentation](https://learn.microsoft.com/en-us/azure/foundry/fine-tuning/overview) puts it in one sentence: fine-tuning adjusts the model’s weights, rather than adding examples to each prompt. The weights are the numbers that hold what the model has learned. With a [prompt](/en/glossary/prompt) or with [RAG](/en/glossary/rag), the model is the same and what changes is what it has in front of it; with fine-tuning, what changes is the model. That’s why fine-tuning is done once, before the model is used, and what it learns stays inside until it’s trained again. What the agent reads with each answer, on the other hand, can change every day without touching anything. ## How does it work? Microsoft distinguishes three approaches, depending on what the model is given to learn from. The most direct is **supervised fine-tuning** (SFT): it’s given example inputs with the right answer next to them, for example customer messages and how the office answered them well. It works when what you want can be taught with examples. The second is **preference fine-tuning** (DPO): for each input it’s given two answers, a preferred one and a rejected one, and it learns to be more like the first. It fits when there’s no single right answer and what matters is tone or completeness. The third is **reinforcement fine-tuning** (RFT): instead of answers, it’s given a grader that scores what it writes, and the model learns to score better. It’s for when there are several valid solutions and you can measure well which is better. The whole model isn’t always retrained. Parameter-efficient techniques, which IBM groups under the acronym PEFT, change only a small part of the parameters and greatly reduce the computing and memory needed. One of them is **LoRA**: the adjustment is stored as a separate piece that plugs into the base model, so you can have several and switch between them depending on the task. ## What is it for? According to Microsoft, fine-tuning helps when there is good, specific training data and you want one of these things. For the model to be more accurate at a specific task, because with more examples than fit in a prompt it learns the patterns. For the prompt to be shorter, because what used to be explained with every answer it now knows, which lowers the cost per token and the response time. For it to always answer with the same tone, format or structure. There are three more technical uses. Improving how it uses tools, choosing the right one and filling in its data correctly. Making better use of the documents it’s given, learning to keep what matters and ignore the rest. And specializing a small model, cheaper and faster, with examples generated by a large one, so it does a specific task almost as well. In a field service company, that translates into things like classifying service requests by type and urgency with the company’s own criteria, always extracting the same fields from an email in the same format, or getting a small, fast model to do a very repetitive task well, like recognizing the brand and error code in a customer’s message. ## What isn’t it for? Making the model know facts that change. It’s the most common misunderstanding, and Microsoft says so in the same paragraph where it introduces it: Fine-tuning documentation on [Microsoft Learn](https://learn.microsoft.com/en-us/azure/foundry/fine-tuning/overview), accessed on November 14, 2026. If rates change in January or the on-call technician changes every week, a model trained on December’s data will keep saying December’s, with total confidence, until it’s trained again. That’s what RAG is for: the agent reads the fact in your system or your documents at the moment it answers. Nor does it replace safety controls: what the agent must never do is controlled through its permissions and its rules, not by hoping it has learned it. ## Fine-tuning, RAG or prompt? All three adapt a model to a company, and you almost always start with the cheapest. The **prompt** comes first: writing clearly what it has to do, what it doesn’t know and how it should answer solves a huge part of the work, and it can be changed in minutes. **RAG** comes second, for everything the agent needs to know about the company that changes: procedures, rates, contracts, each unit’s history. **Fine-tuning** comes later, when the prompt and the documents are already right and there’s still something the model doesn’t do as you want: a format it doesn’t stick to, a classification it always gets wrong in the same way, or a cost per answer that only comes down with a smaller model trained for that task. The point-by-point comparison between RAG and fine-tuning is in [what is RAG](/en/glossary/rag). ## What do you need to do it? First, **good examples**. Microsoft insists they be correct, consistent and representative, and that you check before uploading them for missing cases, duplicates, bias or sensitive information. In field service, the examples are conversations with customers, so they contain names, addresses and phone numbers: that’s personal data, and before using it for training you need to be clear about consent and remove what isn’t needed. Second, **a way to measure**. The training examples have to be kept apart from the test ones, and the tuned model compared with the starting model on cases it hasn’t seen, not just by watching the error go down during training. Microsoft calls it an iterative engineering process, not a guaranteed improvement. And third, **budget and maintenance**. You pay for training, for repeated tests and, often, for hosting the tuned model. And when the base model is retired or the data changes, it has to be retrained. At [Delegic](/en), that’s why what each company’s agent knows lives in its instructions and its documents, which the company can read and correct, and not inside a trained model nobody can open. ## Frequently asked questions ### What does fine-tuning mean? Continuing to train an existing model, with your own examples, to refine it for a task. ### Can I train ChatGPT on my company’s data? The big providers let you fine-tune some of their models through their business platforms. But if what you want is for it to answer with the company’s data, the usual and cheaper way is RAG: giving it the documents at the moment it answers, without training anything. ### How many examples do you need? It depends on the task and the model, and there’s no figure that fits everything. What the guides do repeat is that quality and variety matter more than quantity, and that some must always be kept back for testing. ### What is LoRA? A parameter-efficient tuning technique that doesn’t change the whole model: it stores the adjustment as a small separate piece that plugs into the base model. It’s cheaper to train and lets you keep several pieces for different tasks. ## Related terms - [LLM](/en/glossary/llm): the model that gets tuned. - [RAG](/en/glossary/rag): giving it the documents with each answer, without training. - [Prompt](/en/glossary/prompt): the instructions the model reads before it answers. - [Evaluation](/en/glossary#evaluation): how you check whether a change makes things better or worse. - [Embeddings](/en/glossary/embeddings): the numbers classic RAG uses to search by similarity.