Back to the blog

AI·7 min read

How to train an AI on your company documents

You almost never need to train it: you connect it to your documents. What to prepare, what happens to your data, how to test it and what to ask first.

RAG
AI on your documents
Searches, isn't trained
Says which document
Respects who sees what
In this article

You don't need to train it. In almost every case the AI is connected to your documents, searches them before answering and tells you which one the answer came from: that's called RAG. Training your own model is mostly for something else, costs more and goes stale with the first new price list.

When a company asks to "train an AI on our documents", it almost always wants it to answer with its product sheets, price lists and procedures, without making anything up. That comes from having the documents in order, testing it before opening it up and a clear contract with whoever builds it.

Do you train the AI or connect it to your documents?

You connect it. Microsoft explains it in its guide to customizing a large language model (updated 27 February 2026): RAG brings external data into the prompt the model receives, and it's the way to answer with a company's private data or with information newer than what the model learned. Fine-tuning, what most people mean by "training", is for Microsoft an advanced capability: its good cases are getting the model to write in a specific style, tone or format, and it asks you to measure how the model does without it first. It needs high-quality examples in a special format, and it has a training cost plus an hourly cost while the custom model runs.

OptionWhat it doesWhen a price list changesIt suits you if...
The tool you already use (ChatGPT, Claude or Copilot)Uses the files you give it as contextIf you uploaded a copy, you upload the new oneFew documents and a small team
An assistant connected to your documents (RAG)Searches them before answering and says which one it usedReads the new one, if it's built properlyLots of documents, or customers asking
Training a model (fine-tuning)Adjusts a model with examplesIt has to be trained againYou want a style or a format, not data that changes

When is the tool you already use enough?

Often, to start with. In Claude you upload documents to a project and it uses them as context in that project's chats, and on paid plans it switches to a RAG mode by itself when the material grows (Claude help center). Microsoft 365 Copilot works with the organization's data that each person already has permission to see (its privacy page, updated 3 November 2026).

Our take: it's enough with a few documents and nobody from outside asking. It falls short when everyone uploads their own copy and nobody knows which one is right, when not everyone can see everything (distributor price lists, say) or when you want customers to ask from your website.

Which documents do you need ready?

The slow part is almost never the AI: it's the documents. Before asking for a quote, take this inventory:

  1. Which documents answer the everyday questions. Product sheets, price lists, terms of sale, lead times, procedures and the reply emails that already work.
  2. One good version of each. With two different price lists, the AI will answer with one of them, and not always the right one.
  3. Who keeps it up to date. A document with no owner goes stale and the assistant keeps answering with it.
  4. Who can see what. Public, team and management material don't go together.
  5. What stays out. Customers' personal data, payroll, contracts with names. If it isn't needed to answer, leave it out.

What happens to your data?

This is what they say about their business versions and their API (checked on 10 November 2026):

  • OpenAI: by default it doesn't use data from ChatGPT Enterprise, Business and Edu or from its API to train its models, and it offers to sign a Data Processing Addendum (business data).
  • Anthropic: it doesn't train on chats from its commercial products, unless the company joins its Development Partner Program or sends material on purpose, for example with thumbs up or down (privacy center, 16 March 2026).
  • Microsoft: prompts, responses and data Copilot accesses through Microsoft Graph aren't used to train foundation models (Microsoft 365 Copilot privacy).

With a personal account, check the settings before uploading anything.

And if whoever builds the assistant processes personal data on the company's behalf, they're a processor under Article 28 of the GDPR: Spain's data protection agency (AEPD) reminds you in its model clauses that the arrangement has to be in writing. Our rule, published on our business page: your customers' personal data never goes into a tool whose contract doesn't allow it.

How do you test it before giving it to the team or to customers?

With real questions and the right answer next to each one. Even the AEPD, in its internal policy on generative AI (12 December 2025), lists hallucinations among the risks and rolls it out with pilot tests, ongoing monitoring and human oversight.

How we do it:

  1. Collect 30 real questions from the team and from customers, with the right answer and the document it comes from.
  2. Add five whose answer isn't in any document.
  3. Note whether it gets it right, whether it says which document it used and whether it says "I don't know" for the five with no answer.
  4. Run the test again every time an important document changes.
Question (made-up example)Right answerGets it right?Says where from?
What's the lead time for the four-shelf wooden display?Three weeks (2026 price list)YesYes
Do you ship to the Canary Islands?Not in the documentsNo: makes up a yesNo

A mistake on prices, lead times or terms ends in a promise you can't keep: if it fails there, it doesn't go live until it's fixed.

What should you ask before hiring someone to build it?

  1. Does it answer only from our documents and say which one each answer comes from?
  2. What does it do when the answer isn't there: say it doesn't know or make it up?
  3. Which model does it use, under what contract, and where are our documents stored?
  4. Does it respect who can see what?
  5. When we update a price list, does it pick it up by itself or does someone have to be told?
  6. How do we test it before opening it up, and who writes the questions?
  7. What do we pay each month for model usage, and whose name is the account in?
  8. Do we sign the data processing agreement?
  9. If we leave, do we take the documents, the instructions and the tests with us?

How do you measure before and after?

With minutes per task, before and after, as we explain in the FUNDAE post. With an assistant, add three numbers:

  • How many questions reach the person who knows everything, the week before launch and four weeks after.
  • How long it takes to answer a customer enquiry.
  • The 30-question test, every month. If the score drops, a document changed and nobody updated it.

When is it not worth it?

  • If the answers live in the head of someone who never wrote them down. Write them down first.
  • If the documents contradict each other and nobody decides which one counts.
  • If the answer is a legal or money decision: there the AI drafts and a person signs.

Is there an example already running?

Boix PLV, a display manufacturer and a client of ours, has an AI product search on its website: the customer types what they need, for example a sturdy display for 50 bottles of wine, and it returns the display from their catalogue that solves it. It answers with what Boix makes, not with what's on the internet. It's in production and has no published figures yet.

In Labs there's a demo with a made-up company and a real collective agreement: it drafts the working-conditions document required by Spain's RD 723/2026, flags what doesn't add up and a professional reviews it.

How much does it cost with us?

We publish it with our work for companies: a case built on your documents starts at €4,500 plus VAT and includes measuring before and after. It starts with a 1 to 2 week diagnosis, which gives you the price and timeline in writing, and you pay by phases. The code and the data are yours. If you'd rather have the team learn to use AI on their own tasks first, the workshop is €1,200 per group plus VAT, eligible for FUNDAE funding.

Frequently asked questions

Can I train ChatGPT on my company documents?

You can fine-tune a model, but you don't need to for it to answer with your documents. The usual way is to connect it to them (RAG): it searches before answering and stays current when they change.

Does the AI use my company documents to train itself?

In the business versions from OpenAI, Anthropic and Microsoft, not by default, according to their own privacy pages.

Can it answer customers or only the team?

Both, with different permissions. For customers, only public material, and only with no mistakes on prices, lead times and terms.

How much does an AI assistant on company documents cost?

With us, from €4,500 plus VAT, with a prior diagnosis and before-and-after measurement. Also ask what you'll pay each month for model usage.


Run the nine questions past whoever quotes you for the assistant, us included. If it adds up, write to us with three or four real questions from your team and we'll see whether what you already have is enough.

Topics

AI in business · AI assistant · RAG · company documents · companies

Share

Fast Web Academy

One emailwhen there's something to tell.

What we test in the lab and what works for clients. Written in Spanish. No spam.