Every month a prospect asks me if they need to fine-tune. 90% of the time the answer is no. Here is how to know when you're in the other 10%.
The four levers
Prompting, RAG, tool use, then fine-tuning. The first three are reversible and cheap — exhaust them first.
When fine-tuning is right
Consistent style, ultra-strict format, latency-critical, massive recurring cost, proprietary domain knowledge.
When it is a mistake
To teach new knowledge (use RAG), to patch occasional errors, without clean eval set.
