Fine-tuning vs RAG vs reinforcement learning: which does your company need?
Three ways to make an AI model work better on your company's tasks, explained without the jargon, in the order you should try them.
The three side by side
| RAG | Fine-tuning | Reinforcement learning | |
|---|---|---|---|
| What it does | Looks up your documents at answer time | Trains the model on examples of good answers | Trains the model by rewarding attempts that succeed |
| Best for | Questions about your policies, products, contracts | A fixed format, tone, or narrow classification | Multi-step work with a checkable result |
| What you need | Your documents | Hundreds to thousands of good examples | Realistic tasks plus a way to score each attempt |
| Effort | Days to weeks | Weeks | Weeks to months |
| Changes the model? | No | Yes | Yes |
RAG: start here
Retrieval-augmented generation means the model searches your documents and answers from what it finds. The ChatGPT and Claude business plans already do a simple version through projects and connectors. It's the right first step for most questions about your own information, and it keeps working when the documents change.
It falls short when the task isn't about finding facts but about doing something your way: a judgement call, a specific output format, a long sequence of steps.
Fine-tuning: when instructions aren't enough
Fine-tuning trains an existing model on examples of input and ideal output. Use it when you've written careful instructions and the model still drifts: the wrong structure, the wrong tone, inconsistent labels. It often lets a smaller, cheaper model do a narrow job as well as a large one.
The catch: you need a good set of examples, and the model only gets as good as them. Fine-tuning teaches style and pattern well; it teaches new facts poorly. Use RAG for facts.
Reinforcement learning: for tasks with a right answer
RL is how the leading AI labs teach models to do long, multi-step work. The model attempts a task many times in a realistic environment, an automatic checker scores each attempt, and the model is trained toward the attempts that scored well.
It fits company work where success can be checked: did the invoices reconcile, did the claim get coded correctly, does the drawing meet the spec. It needs three things most companies don't have ready: a realistic environment to practise in, a set of real tasks, and a checker that can't be fooled.
This is the work we organise through owned models: we build the benchmark from your real tasks, and specialist labs bid to train the model, paid when it passes.
The order to try them in
- Better instructions and examples in the prompt.
- RAG over your documents.
- Fine-tuning for format, tone and narrow skills.
- RL for multi-step tasks with a checkable result.
Questions
Is fine-tuning better than RAG?
They solve different problems. RAG gives a model access to facts it doesn't know; fine-tuning changes how it behaves. Many systems use both.
Can I fine-tune ChatGPT or Claude?
OpenAI offers fine-tuning for some of its models through its API, and open-weight models can be fine-tuned by anyone with the hardware. Check each provider's current options.
What is reinforcement learning in AI, simply?
The model tries a task many times, each attempt is scored automatically, and the model is trained toward the attempts that scored well.