In one sentence

Fine-tuning is adjusting a model by training it further on your own examples, so it behaves more the way you want on a specific task.

The simple model

A general model is trained on broad data. Fine-tuning takes that model and trains it a bit more on examples of your particular task, so it gets better at that narrow thing, a specific format, a consistent style, a specialized classification.

It sounds like the obvious way to make a model "know your company." For most workflows, it is not the first thing to reach for.

Why you usually do not need it first

Two simpler approaches solve most needs, and you should try them before fine-tuning:

  • For giving the model your current, company-specific information, use retrieval. It lets the model answer from your documents without any training, and stays current as the documents change. Fine-tuning does not update knowledge; retraining does.
  • For shaping how the model behaves, a clear prompt and good examples in the request often get you most of the way.

Fine-tuning adds real cost and effort: gathering quality examples, running the training, and repeating it when things change. That effort is only worth it once the simpler options are genuinely not enough. This is use the least complex mechanism applied to models.

When fine-tuning can help

It becomes worth considering when you need consistent behavior on a narrow, stable task that prompting cannot reliably produce, and you have a good set of examples to train on. Even then, it is a step to take deliberately, not by default.

A small example

A team wanted an assistant to answer from their handbook and considered fine-tuning a model on it. Retrieval solved it instead: the assistant answered from the current handbook and updated automatically when the handbook changed. Fine-tuning would have frozen the knowledge and needed redoing at every update.

What this does not solve

Fine-tuning can shape behavior; it does not give a model current knowledge, guarantee correctness, or remove the need for guardrails and review. It is one advanced option among several, useful in specific cases, and rarely the right starting point.