You want the chatbot to know your returns policy, your product names and the way your company writes. Somewhere you read that the answer is to “train it on your own data”, and now you are looking at three terms that all sound like engineering projects: fine-tuning, RAG and prompting.
For most people asking, the answer is that you do not need to train anything. You need to put the right material in front of the model, and ChatGPT, Claude and Gemini each have a feature for that. Fine-tuning is a real tool for a narrow set of problems, and the best-known self-serve version of it is being wound down. OpenAI stopped taking on new fine-tuning organisations in May and ends new training jobs for its remaining customers on 6 January 2027.
This comparison is built from OpenAI’s and Google’s own documentation, checked on 27 September 2026. Nothing here was tested hands-on. Where we are drawing a conclusion rather than reporting one, we say so.
The short version
- Prompting changes what you tell the model this time. It is the cheapest and fastest option, and the vendors’ own first recommendation.
- RAG, or retrieval, searches your documents and pastes the relevant passages into the prompt. It fixes “it doesn’t know our stuff”.
- Fine-tuning retrains the model on examples of good answers. It fixes “it knows what to do but won’t do it consistently”, usually for a format or a label, at volume.
- If your problem is missing facts, fine-tuning is the wrong lever.
What each one actually changes
All three change what comes out of the model. They differ in where they step in.
Prompting changes the input. Instructions, an example of good output, and any material the model needs all go into the message you send. OpenAI’s model optimisation guide tells developers to “include text or image content that the model will need to generate a response from outside its training data”, and says prompt engineering “may be all you need in order to get great results for your use case.”
Retrieval also changes the input, but automatically. Your documents are cut into chunks and indexed. When you ask a question, a search step picks the chunks that look relevant and places them in the prompt. We have covered what uploading a document actually does, including why the search sometimes misses the passage you needed. The point here is that retrieval is still prompting. The model is handed text and answers from it.
Fine-tuning changes the model. You supply examples of prompts paired with ideal responses, and training adjusts the model’s internal weights so it produces that kind of response by default. OpenAI’s fine-tuning guide describes the result as “a customized model that more reliably produces your desired style and content.”
That difference between changing the input and changing the weights decides almost everything else.
Match the symptom to the fix
Start from what is going wrong, not from the technique.
| What’s going wrong | Likely lever | Try this first |
|---|---|---|
| It doesn’t know your products, policies or recent events | Retrieval, or simply pasting the material in | Attach the document, or add it to a project |
| It knows the material but answers vaguely or in the wrong shape | Prompting | Say exactly what you want and show one example |
| It drifts from your instructions in long conversations | Prompting and context | Start a fresh chat with the key instructions at the top |
| It invents details that aren’t in your document | Prompting and retrieval | Tell it to answer only from the attached text and to say when the answer isn’t there |
| Output format drifts across thousands of automated runs | Fine-tuning, eventually | Tighten the prompt, add examples, measure, then consider tuning |
| A specialised classification or translation task at volume | Fine-tuning | Collect at least 50 examples of correct answers before anything else |
The first four rows cover most of what people mean by “make the AI know my business”. None of them involves training.
Row four deserves a sentence of its own. A model that invents details is doing what language models do when the answer isn’t in front of them. Giving it the source, plus permission to say “not stated”, fixes more of that than training would.
What the vendors themselves recommend
OpenAI’s and Google’s documentation give the same order of operations.
Google’s tuning documentation says: “We recommend starting with prompting to find the optimal prompt. Then, move on to fine-tuning (if required)”. It describes fine-tuning as most effective when you have “a sizable dataset to train on (think 100 examples or more)” and for tasks “where advanced prompting strategies are not sufficient.”
OpenAI’s fine-tuning guide lists what supervised fine-tuning is best for: classification, nuanced translation, generating content in a specific format, and correcting instruction-following failures. It sets a minimum of 10 examples, recommends starting with 50 well-crafted ones, and includes a line worth keeping: “If 50 examples have no impact, rethink your task or prompt before adding training data.” It also tells developers to set up evaluations, meaning a repeatable way to measure whether the output improved, before investing in fine-tuning at all.
Look at that “best for” list again. Nothing on it is “knowing your returns policy”. That is our reading rather than a vendor statement, but the reasoning is simple. A fact that lives in a document can be updated the day the policy changes and quoted back with a page reference. A fact trained into the model’s weights can do neither.
OpenAI is winding down self-serve fine-tuning
Many guides that present fine-tuning as the natural third step were written before this. OpenAI’s deprecations page sets out the timeline:
| Date | What changes |
|---|---|
| 7 May 2026 | Creating fine-tuning jobs no longer available to organisations that had never run fine-tuning |
| 2 July 2026 | Also closed to organisations that had not run inference on a fine-tuned model in the past 60 days |
| 6 January 2027 | Active existing customers can no longer create new fine-tuning jobs |
Models that have already been fine-tuned keep working. OpenAI says inference on them continues until the underlying base model is deprecated. Its fine-tuning guide now carries a notice that the platform “is no longer accessible to new users”, and the optimisation guide says its fine-tuning material is being moved into legacy documentation.
OpenAI’s pages do not give a reason, and we are not going to invent one. What it means in practice is plain: if you are not already an OpenAI fine-tuning customer, it is not an option there. Google still documents tuning for Gemini models on its enterprise cloud platform, so the technique has not gone away. It has moved further from anything a non-developer would set up.
When fine-tuning still earns its place
It is the right tool when all of these are true:
- The task runs many times, automatically, rather than in a chat window.
- You can say what a correct answer looks like, and you have dozens to hundreds of real examples of it.
- You have already tried a careful prompt with examples, measured the result, and it still falls short.
- The problem is behaviour, such as format, labels or consistency, not missing knowledge.
Google also lists lower latency and lower cost as benefits, because a tuned model needs shorter prompts. Once a behaviour is trained in, you stop paying to send long instructions with every request. That matters at volume and is irrelevant to someone typing into a chat app. If per-request cost is your concern, here is what an AI bill is actually made of.
What to do this week
If you want an assistant that knows your business, work through these in order.
- Write the instructions once, properly. Who it is writing for, the tone, and what to do when it doesn’t know. Our practical guide to prompting covers the parts that matter.
- Give it the documents in a place that persists. ChatGPT and Claude call this a project, and Gemini calls its version a Gem. All three appear on the vendors’ free plans; Claude’s free plan allows up to five projects.
- Tell it to stay inside the documents. “Answer only from the attached files. If the answer isn’t there, say so.” Then check a few answers against the source yourself.
- Keep the material to what’s needed. Past a certain size, these features search your files rather than reading them whole, which is where retrieval misses begin. Here is how much fits in one go.
Fine-tuning is worth a conversation only if you are running the same task at volume, have worked through steps one to three, and have measured the shortfall. That conversation will be with a developer, not a settings menu.