The instinct and the reality
Teams reach for training when output is not quite right, imagining the model needs to learn their business. Usually the model has not been given the right information or clear enough instructions, which are both far cheaper to fix.
Try context first
Most “it doesn’t know our business” problems are retrieval problems: the relevant document was not supplied, or the corpus contains contradictions. Fixing that is a scoping exercise, not a training run.
Then try examples
Showing several examples of exactly the output you want moves quality more than most people expect, particularly for format and tone — the same reason voice comes from examples rather than adjectives.
“The model doesn’t know our business” almost always means “we didn’t give it our business.”
Then constrain the output
If the problem is inconsistent structure, requiring a defined shape solves it deterministically. That is the strongest guardrail available and it needs no training at all.
When fine-tuning genuinely helps
A narrow, repetitive task at high volume where you have many labelled examples and need lower latency or cost than prompting achieves. That is a real case — and it is a much smaller set of situations than the interest in it suggests.
The cost is the commitment
A tuned model is a version you own, maintain, re-tune as data drifts, and re-do when you want to move to a better base model. It converts a swappable dependency into a sticky one, which is the real price.