Fine-tuning is the process of taking an AI model that has already been trained and training it a little further on a smaller set of your own examples, so it becomes better at one particular task, format or style. The result is a new version of the model that behaves differently by default, without needing long instructions in every request.
How fine-tuning works
A foundation model learns general language skills from a huge body of text. Fine-tuning starts from that model and shows it hundreds or thousands of paired examples: an input and the ideal output. Each pass nudges the model’s internal weights towards producing outputs like yours.
A marketing example: a retailer with 5,000 product descriptions written in its house style could fine-tune a model on pairs of “supplier specification in, finished description out”. Afterwards, the model writes new descriptions in that style from a specification alone.
Most businesses do not fine-tune from scratch. At the time of writing (October 2026), several AI providers offer fine-tuning through their developer platforms for some of their models: you upload a file of examples, the provider runs the training, and you get a private model to call. Lighter methods adjust only a small part of the model, which makes the process cheaper and faster.
Fine-tuning is one of three ways to shape a model’s output. The other two are usually cheaper and should be tried first:
- Better prompting Including few-shot prompting, where examples go into the prompt rather than into the model.
- Retrieval-augmented generation Where the system looks up relevant documents and passes them to the model at the time of the question. This is the right tool when the problem is missing or changing facts.
As a rule, fine-tuning teaches a model how to respond (format, tone, labels), while retrieval gives it what to say (current facts). Fine-tuning a model on your price list is a poor idea: the prices will change and the model will keep repeating the old ones.
Why it matters
For most UK small businesses, fine-tuning is not the first step and may never be needed. It earns its cost when you have a high-volume, repetitive task, a large set of good examples, and prompting has hit a ceiling. Examples include classifying thousands of customer reviews, producing product copy at scale in a strict format, or routing support emails.
There is a data protection side too. Training or fine-tuning a model on customer data needs a lawful basis and transparency under UK GDPR. If your examples contain names, emails or chat transcripts, you need to have told people their data may be used this way, consider whether the purpose is compatible with why you collected it, and apply data minimisation by stripping out personal details the task does not need. Personal data used in training can be hard to remove from a model later.
Common mistakes
- Fine-tuning to add knowledge. Facts change; retrieval handles them better.
- Training on mediocre examples. The model learns your average, not your best. Curate the training set.
- Too few examples. A few dozen examples rarely beats a well-written prompt.
- No evaluation set. Without held-back examples to test against, you cannot tell whether the tuned model is better.
- Uploading raw customer records. Anonymise first, and check the provider’s data processing terms.
How to act on it
Write the task down and try to solve it with a strong prompt and good examples first. Measure the result on 20 to 50 test cases. If that falls short and the task is frequent, gather a clean set of your best examples, remove personal data, and keep a portion back for testing. Compare the tuned model against your prompt on the held-back set before switching over, and record what data was used and on what lawful basis.
Deciding where AI belongs in your marketing, and where it does not, is a question I cover in digital marketing strategy and consulting, starting from the tasks that cost you the most time.