Working with AI
What is Fine-tuning?
Fine-tuning is additional training that adapts an existing AI model to your specific task or style using your own examples, changing the model’s weights rather than just its instructions.
Where prompting tells the model what to do at request time, fine-tuning teaches it beforehand. You provide hundreds or thousands of input-output examples of exactly the behaviour you want (your tone, your classification labels, your document formats) and train a variant of the model that behaves that way by default. It shines when the task is narrow, repetitive, and needs consistency: always classify into these 14 categories, always write in this house style, always parse this odd legacy format.
It is also the most oversold technique in AI sales. The blunt sequencing: prompt engineering first, RAG for knowledge, and fine-tuning only when both plateau. Fine-tuning does not give the model new facts about your business (RAG does that better and stays current), it costs real money to build the training set, and it locks you to a model version you now maintain. For knowledge-heavy problems ("answer from our policies") it is usually the wrong tool bought for the right instinct.
Where it earns its keep: shrinking costs at volume. A fine-tuned small model can match a large model on one narrow task at a fraction of the per-call price, which matters when you process a hundred thousand items a month. That is the pattern we reach for in cost-sensitive pipelines: large models to design and evaluate, small tuned or tightly-prompted models to run the volume. Ask any vendor proposing fine-tuning one question: what did prompting plus retrieval fail to do? If they have no specific answer, they are selling the expensive option first.
Frequently asked
- When should I use fine-tuning instead of RAG?
- Use RAG when the goal is knowledge, answering from your policies and current data, because it stays up to date and fine-tuning does not reliably add facts. Reach for fine-tuning when the task is narrow and repetitive and needs consistency: always classify into these 14 categories, always write in this house style. The honest sequence is prompting first, RAG for knowledge, fine-tuning only when both plateau.
- Is fine-tuning expensive?
- Building the training set costs real money and time, and you then maintain a model version, so it is the most oversold technique in AI sales. Where it earns its keep is shrinking cost at volume: a fine-tuned small model can match a large one on a single narrow task at a fraction of the per-call price, which matters when you process a hundred thousand items a month.
- How do I know if a vendor is proposing fine-tuning for the right reason?
- Ask one question: what did prompting plus retrieval fail to do? A real answer names a specific limit they hit. If they have no specific answer, they are selling the expensive option first.
Wondering where this fits in your business?
The free AI audit maps your operations and shows the three systems we would build first, in plain language like this.
