Fine-tuning is extra training on examples you wrote: a few thousand replies, done your team’s way. The model keeps everything it knows. What moves is tone and habit. Same brain, different manners.
Manners, not facts
Fine-tuning changes how the model answers: your tone, your format, your categories. It files each question the way your team already does.
What does not change is what the model knows. Three thousand replies teach it how the bakery writes, nothing new about bread. Facts go in front of the model instead.
Hear the difference
One inbox, three questions, two versions of the same model. The fine-tuned one read three thousand of this bakery’s replies before its first shift.
When to skip it
Most jobs need a better prompt. Write the perfect reply once, paste it in as an example, and the model copies it. You are finished.
Fine-tuning pays when the model still drifts by the hundredth message. It takes a few thousand examples, a checker, and a budget. Pay for drift you can measure.
Working on this?
Not sure whether your job needs a custom model? Tell me what it is. I read every one of these.