If you're wondering how to train AI on your own data, the short answer is that you almost never need to train it. For it to answer with your company's facts, it's enough to give it your documents to read: it looks them up before answering and, when a fact changes, you change the document. Retraining the model is a different job, technical and expensive, that a small business rarely needs.
The example in this guide is a made-up artisan creamery that sells to shops and restaurants.
A language model is the program behind ChatGPT and every other assistant: it learned to write by reading enormous amounts of text (we cover it in what an LLM is). Training it means changing it from the inside so it behaves differently. That's an engineering job.
People who search for how to train AI on company data almost always want something else: answers that use their rates, their products and their rules. You get that by putting the paperwork in front of it. There are three paths, and only the third one is training.
It's what you already do without a name for it: you copy some text or attach a file and ask. What you type in that box is called a prompt: the request plus everything that goes with it.
When it works: for a one-off question with a few papers. The owner of Three Gables Creamery attaches the labeling rules and asks whether her new label meets them.
Its limits: you have to repeat it in every conversation, each person pastes their own version, and only so much text fits. For five PDFs it's fine. For the whole company it falls short.
Here the documents are uploaded once to a shared place, and the assistant searches them every time somebody asks. The technique is called RAG, “retrieval-augmented generation”: before writing, it retrieves the pieces of your documents that apply and answers with them in front of it. The set of documents is called a knowledge base.
OpenAI frames it the same way in its developer guide: when the model is missing information, or has it out of date, you retrieve it from your documents. Retraining is for something else.
This one is training. It's called fine-tuning: you give the model many examples of a question and a well-done answer so it learns a behavior, like a very rigid format or a way of reasoning.
Per that same guide, it fixes problems of behavior and consistency, not of knowledge. You have to prepare the examples, have someone technical on hand, and do it again when something changes. And a retrained model can't show which document a fact comes from, because it doesn't come from any.
When it makes sense: in a program that does the same task thousands of times and needs an exact format. In a small business that wants its assistant to know the place, almost never.
If what you were looking for was how to train ChatGPT on your own data, in practice that means uploading your files to it. There are two ways:
We don't give figures because they change often; today's are in OpenAI's help article on uploading files and in the one on projects. The part about when to retrieve and when to retrain comes from its guide to model accuracy, checked on November 12, 2026.
If the question is for today and won't come up again, stay on path 1. And if almost nothing at your company is written down, start by writing the essentials: without that, whatever you set up is an empty folder with a search box.
| Path | What you need | How long it takes | When a fact changes | Does it show the source? |
|---|---|---|---|---|
| 1. Paste it into the conversation | Copy and paste, or attach the file | Minutes | You paste it again | No need: it's right in front of you |
| 2. Documents it looks up (RAG) | The documents gathered, and someone to maintain them | An afternoon to get started | You change the document | Depends on the tool. Insist on it |
| 3. Retrain the model (fine-tuning) | Many prepared examples and someone technical | A project, not an afternoon | You have to train again | No: the answer doesn't come from any document |
Times are rough guides, for a small business that already has its documents written.
Three situations at the creamery. Pick the one closest to yours.
If you have five PDFs and a one-off question
If you have a hundred documents that change every month
If you want it to write exactly in your style
Five steps for path 2. Get these right and the rest is clicking a button.
The price list, the product sheets, the shipping terms, the emails you send over and over. Don't write anything new yet.
What you do and for whom, who your typical customer is, what you sell, how you talk, and what can't be said.
Last year's price list and the promotion that already ended only confuse things. If you're unsure about a document, leave it out.
In current formats: PDF, Word, Excel. Scanned pages work if the tool transcribes them.
Ask it the ten questions shops ask you most. Wherever it fails, a document is missing or an old one needs to go.
In Flipa, path 2 is uploading a document. Here a PDF of frequently asked questions goes up, Flipa reads it and files it, and the meter for what it knows about the company moves from 65% to 67%. The animation recreates the real screen of a test account, with the wait shortened. The app's interface is in Spanish today, and English isn't available yet; the assistant answers in the language you write in.
You almost never train it: you give it your data to read. You upload your documents to a tool that looks them up before answering and, when a fact changes, you change the document. Retraining the model is a technical job that's good for fixing a behavior, not for teaching it about your company.
For the first two paths, no. Pasting text or uploading files happens on screen. Retraining a model is technical work (preparing many examples, then running and checking the result), and a small business almost never needs it.
In practice, by uploading files. In a conversation, use the “+” button in the message box. To keep them available, put them in a project, which stores files and instructions for all of its conversations. How many files fit depends on the plan and changes often: check OpenAI's help center.
A few good ones. Five topics told well (what you do and for whom, your typical customer, what you sell, how you talk, and what can't be said) cover the main questions. In a Flipa test with a made-up company, those five took the knowledge meter to 64%.
With documents it looks up, you replace the document and the next answer uses the new one. In Flipa you can also tell it in the chat: it proposes the change with the before and after, and saves nothing until you approve it. That works for notes and brand documents in text, not for a PDF. With a retrained model, you'd have to train it again.
Once you have it, start free and give it to Flipa to read.
Start free