Fine-tuning vs Prompting.

There are two main ways to adapt an LLM to your specific task: prompting (giving the model good instructions for each request) and fine-tuning (retraining the model on examples of your task). For most SMB applications, prompting is the right answer. Fine-tuning is needed in a small minority of cases.

Prompting works because modern LLMs are big enough and capable enough to follow detailed instructions reliably. You don't need to train the model to extract data from your specific invoices — you just need to tell it what fields you want, give it some examples, and define the output format. This is called few-shot prompting and it handles 90% of business use cases.

Fine-tuning makes sense when: the task is very specific to your domain (specialist medical or legal language), you need to consistently match a very particular output style, you're processing such high volumes that the cost of long prompts becomes meaningful, or you have specific data that the base model genuinely doesn't know.

For 95% of SMB use cases — document extraction, enquiry triage, quote generation, internal assistants — well-engineered prompts with good context (often via RAG) are faster, cheaper, and just as accurate as fine-tuning. Fine-tuning takes weeks, costs thousands, and locks you to a specific model version. Prompting takes hours, costs almost nothing, and lets you switch between Anthropic, OpenAI, and Google as the models improve.

If an AI consultancy proposes fine-tuning for your invoice matching, ask them why prompting wouldn't work first.