AI glossary · Building and connecting
Fine-tuning
Also called: Fine tuning · Further training
What is fine-tuning?
A language model first learns from huge amounts of text to predict the next word. It is then trained further so it is useful as an assistant. According to Anthropic, Claude has already been fine-tuned this way.
Fine-tuning for a business takes the training one round further, with your own examples. Think of a qualified tradesperson joining a new company. They already know the craft; now they learn how this company writes its quotes.
How does it work?
Fine-tuning runs in three steps:
- Collect examples: many pairs of input and the answer you want, such as a customer enquiry and a reply in your company's tone.
- Train: The model is trained further on these examples and picks up their patterns, mistakes included.
- Test: The new model is tested on cases that were not part of the training.
Why does it matter for your business?
Fine-tuning sounds like “an AI that knows my business”. For knowledge it is the wrong tool: prices, dates and project status change, and a trained model would need retraining every time. That is what RAG is for: the AI looks things up in your current documents.
For tone and format, a good prompt with a few examples is usually enough. It needs no training run and can be changed at any time. Fine-tuning pays off more when there are very many similar tasks and plenty of good examples.
How does OsAI use it?
OsAI projects start with the simplest approach that holds up: a clear instruction in the system prompt, examples from your business and, where your own knowledge is needed, RAG. This works with any major model and stays replaceable.
Whether a special route such as fine-tuning is needed for your task is something OsAI works out in AI consulting.
Frequently asked questions
Can I fine-tune Claude on my business?
Not through the Claude API at the moment: according to Anthropic it does not offer fine-tuning (as of 13 November 2026). Anyone interested is asked to contact Anthropic. For most business tasks, a good system prompt, examples and RAG are enough.
Which is better, fine-tuning or RAG?
For knowledge, RAG is usually better; for style and fixed formats, fine-tuning can help. RAG pulls current facts from your documents and names the source. Fine-tuning changes how the model writes but does not keep facts up to date.
Is fine-tuning worth it for a small business?
Rarely. It needs many clean examples, a training run and testing, and when you switch to a new model you start over. With a prompt, examples and RAG, a small business usually gets there faster and for less money.
Further reading on the blog (in German)
Matching services
Sources
Last updated: