Fine-tuning takes a pre-trained model and continues training it on a curated set of examples for a specific purpose, such as classifying support tickets, writing in a house style or extracting fields from documents. It adjusts the model's weights, unlike prompting or retrieval, which only change what the model sees. Fine-tuning is useful for consistent behavior but is a weak way to teach a model facts that change often.
Fine-tuning versus retrieval versus prompting
These three approaches get confused. Each fits a different problem.
- Prompting: give instructions and examples in the request. Cheapest and fastest to change.
- Retrieval (RAG): add relevant documents at answer time. Best for facts that change or that need citations.
- Fine-tuning: train on many examples. Best for consistent format, tone or narrow tasks at scale.
When SaaS teams use it
Marketing and product teams sometimes fine-tune smaller models for repetitive jobs: classifying search queries by intent, tagging content, generating structured product descriptions in a fixed format, or routing support tickets. Vendors such as OpenAI and Google document fine-tuning options for some of their models.
Fine-tuning does not affect how public AI assistants like ChatGPT or Gemini talk about your brand. Those models are trained by their vendors; you influence them through the public web, not by fine-tuning your own copy.
Example
Classifying thousands of queries
Kettleward, a fictional HR software company, wants every Search Console query tagged by intent and funnel stage. Prompting a large model works but is slow and inconsistent across tens of thousands of rows. The team labels 1,000 queries by hand and fine-tunes a smaller model, which then tags the rest in a consistent format. Spot checks catch the remaining errors.
Common mistakes
Fine-tuning to teach facts
Prices, features and docs change. Retrieval keeps answers current; a fine-tuned model would need retraining every time.
Skipping evaluation
Hold back a test set and compare the fine-tuned model with a well-prompted base model. Fine-tuning is not always better.
Questions
Costs depend on model size, data volume and vendor. For small models and modest datasets it can be affordable, but data preparation is often the larger cost.
It can improve behavior on a narrow task, but it is not a reliable fix for factual errors. Grounding in retrieved sources is the usual approach for that.
Sources
