The problem it solves
When people want an AI model to do something specific, “just train it on our data” is the phrase you hear, and it is usually the wrong instinct. It treats three very different things, with wildly different cost and difficulty, as if they were one. Understanding the distinction is what lets you reach for the cheap, fast option first and the expensive one only when you truly need it.
How it works
Think of the three as a ladder, from the enormous and rare to the everyday.
Training (from scratch) is building the model. This is the process that creates a large language model in the first place: feeding it an enormous amount of text and adjusting billions of internal values until it has learned the patterns of language. It is spectacularly expensive, takes specialized teams and vast computing power, and is done by a handful of major labs. Almost nobody outside those labs trains a model from scratch, and you will almost certainly never need to.
Fine-tuning is adapting an existing model. Instead of starting over, you take a model that is already trained and adjust it further on a smaller, focused set of examples, to shift its behavior, style, or format. Fine-tuning is far cheaper than training from scratch, but it is still a real project: it needs quality data, evaluation, and ongoing upkeep every time the base model changes. Crucially, fine-tuning is for changing how a model behaves, not for adding facts it should know.
Prompting is guiding a finished model. This is what almost everyone actually does. You leave the model exactly as it is and shape its output through what you put in the prompt: clear instructions, examples, and relevant information. It is free, instant, and reversible. You can change the behavior on the next message, with no training, no data pipeline, and no cost beyond the request itself.
The practical wisdom follows directly from the ladder. Start with prompting. If the issue is that the model lacks specific knowledge, add that knowledge with retrieval (RAG) rather than trying to bake it in. Reach for fine-tuning only when prompting and retrieval genuinely cannot deliver the consistent behavior you need, and you have a clear, measurable reason. Most teams that jump straight to fine-tuning discover, after real expense, that a good prompt or a retrieval setup would have solved the problem.
A concrete example
Say you want an AI assistant that answers in your company’s formal brand voice.
The instinct is “fine-tune it on our writing.” But start with prompting: give the model a clear instruction and a couple of examples of your tone, and it may match the voice immediately, for free. If you also need it to answer from your internal documents, that is a knowledge need, so you use retrieval to supply those documents. Only if the tone still is not reliable enough at scale, after prompting has genuinely been exhausted, does fine-tuning become the justified next step. Training a model from scratch never enters the picture.
How it connects
All three describe how an LLM gets its behavior. Prompting is the everyday one, the skill almost every AI user relies on, and it is a full topic in its own right. When the real need is fresh or private knowledge rather than changed behavior, RAG is usually the better answer than fine-tuning.
Knowing which of the three a problem calls for is a core piece of judgment. An AI Solutions Architect and an AI Consultant make exactly this call, and steering a team away from unnecessary fine-tuning is often the most valuable advice they give.