When an AI feature isn't good enough, teams reach for one of three levers: rewrite the prompt, add retrieval, or fine-tune a model. They fix different problems. Picking the wrong one is how a team spends two months fine-tuning a model to "know" a refund policy that changes every quarter.
This lesson gives you a way to choose, with the numbers that usually decide it.
Three levers, three different problems
| Lever | What it changes | Fixes | Doesn't fix |
|---|---|---|---|
| Prompting | The instructions and examples sent with each call | Unclear task, wrong format, missing rules, most quality problems | Knowledge the model doesn't have and you can't fit in the prompt |
| RAG (retrieval) | Which facts are in the prompt for this question | Missing or changing knowledge: your menus, policies, orders, documents | How the model behaves, its tone or format |
| Fine-tuning | The model's weights, by training on your examples | Consistent behaviour on a narrow task, format and tone, doing the job with a smaller, cheaper model | Facts that change, permissions, citations |
The short version: RAG is for knowledge, fine-tuning is for behaviour, and prompting comes first for both.