Fine-tuning sounds like the serious option. But for most internal assistants and document workflows, searching your own documents is faster, cheaper and easier to correct.
About two-thirds of the AI enquiries we receive begin with a request to fine-tune a model. In most of those conversations, fine-tuning turns out to be the wrong tool.
What fine-tuning is good at
Fine-tuning changes how a model behaves: its tone, its output format, how it frames a task and how consistently it structures answers. It teaches a model how to respond. It is genuinely useful when you need a particular style of output at high volume, or when you want a smaller, cheaper model to handle a narrow task well.
What it is bad at
Fine-tuning is a poor way to teach a model facts. Your product catalogue changes every week and your policies are revised regularly. Building that information into the model means retraining it every time something changes, and you cannot easily tell where an answer came from.
The retrieval alternative
Retrieval-augmented generation (RAG) keeps your documents in a searchable index. The model receives the relevant passages together with the question, and answers from them. In a business setting, this gives you three important advantages:
- Updates are immediate: change the document, and the next answer reflects it.
- Answers can cite their sources, which turns an unverifiable claim into one you can check.
- Access control works, because the index can be filtered by who is asking.
How we decide
We start with retrieval and a well-structured prompt. If testing shows that the problems are about format or consistency rather than facts, we then consider fine-tuning. Following that order has saved several clients a six-figure training bill for a problem that a good search index solved.