Fine-tuning changes how a model behaves, and barely what it knows. That one sentence clears up most of the confusion. If you want an assistant that knows your prices, products or manuals, you do not retrain the model: you hand it that information with every question. That is retrieval-augmented generation, the approach behind a custom ChatGPT for business, and it is cheaper, updates instantly and can show the source of each answer.
Fine-tuning predates ChatGPT. In 2018 Jeremy Howard and Sebastian Ruder showed that a model trained on general text could be steered to a specific task with relatively few extra examples. The chatbots themselves are built that way: a base model further trained on conversations and human ratings. Vendors then opened the option to businesses, but it is shrinking: OpenAI is winding down its fine-tuning platform, and new models with good instructions often get just as far.
Prompt, RAG or fine-tuning
| A good prompt | RAG | Fine-tuning | |
|---|---|---|---|
| What it changes | What the model does in this request | What the model knows for this request | How the model always behaves |
| Updating | Instant | Replace a document | Train again |
| Shows a source | Yes, if you include it | Yes, per answer | No |
| Getting started | Almost nothing | Days of work | Weeks, hundreds of examples |
The order is nearly always the same: a better prompt with a few examples first, RAG once the model needs your own information, and training only when both fall short. Writing a good instruction is covered under prompt.
Where the knowledge lives, and what updating costs
Promptknowledge in the question
Getting startedlight
Updatinglight
Shows its sourceyes
RAGknowledge in your documents
Getting startedmedium
Updatinglight
Shows its sourceyes
Fine-tuningknowledge in the model
Getting startedheavy
Updatingheavy
Shows its sourceno
Our judgement. Updating RAG means replacing a document; updating a fine-tuned model means training again.
What a training run really costs
The training run is the cheap part. The real cost sits elsewhere. Someone has to write and check those five hundred examples, and do it again whenever your offer changes. You are tied to an older model while newer generations get better and cheaper. And a fine-tuned model cannot show where an answer came from, which makes AI hallucinations harder to catch. To estimate what a model costs per conversation, use the AI cost calculator.
When fine-tuning does pay
- One fixed format at scaleThousands of product descriptions a month in exactly the same structure, or output that must always fit a strict schema.
- A small, fast taskClassifying emails into ten categories or reading invoices, where a small trained model matches a large one at a fraction of the cost per item; see invoice automation.
- A voice examples cannot captureRarely needed now: current models follow a style well with a few examples in the prompt.
For what most companies want, an assistant that knows their own information, RAG is the better route. How language models work is explained under LLM, how to run such a project in implementing AI in business, and the cost of a full assistant in AI chatbot cost. Reading invoices with AI is its own topic: invoice automation.
What companies ask before training a model
Can I train ChatGPT on my company?
You can build a custom GPT or project with your documents, or have a system built that searches them. That is not fine-tuning, but it does what most companies mean.
How many examples do I need?
Vendors name a few dozen as a minimum, but for a noticeable and stable result it is usually hundreds, all correct and consistent.
Is my data private when I fine-tune?
On the business APIs of the major vendors your training data is not used for other customers. Read the terms all the same.