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 many SMB applications, prompting is the right place to start. Fine-tuning is reserved for cases where a measured evaluation shows that prompting is not enough.

Prompting works because modern LLMs can follow detailed instructions and examples. You may not need to train a model to extract data from your invoices; you can define the fields, provide representative examples and require a structured output. This is called few-shot prompting and it handles many common business workflows.

Fine-tuning makes sense when the task uses unusually specialist language, you need a consistently distinctive output style, very high volume makes long prompts uneconomic, or evaluation shows a repeatable gap the base model cannot close.

For document extraction, enquiry triage, quote support and internal assistants, well-engineered prompts with good context (often via RAG) are usually faster and easier to test before considering fine-tuning. Fine-tuning adds data preparation, evaluation, cost and model-version maintenance.

If an AI consultancy proposes fine-tuning for invoice matching, ask what evaluation showed prompting would not work first.