AI trained on your company's documents: what it really means and what to demand

"We want an AI trained on our documents" is one of the phrases we hear most. The intention is clear and good: an assistant that knows the contracts, the procedures and the company's history, instead of giving generic answers from the internet.

The way to get there, however, almost never involves training anything. And it is worth knowing before you sign, because it changes the price, the timeline and — above all — the guarantees you can demand.

Training and querying are not the same thing

Training a model means modifying the model itself so it absorbs your information. It is expensive, slow, has to be repeated every time a document changes, and has an unpleasant side effect: what it learned is mixed in with the rest of the model's knowledge and can no longer be separated or deleted easily. If a client exercises their right to erasure tomorrow, you have a problem.

What is actually done in practice is indexing and querying: documents are stored in your own repository, and when someone asks a question the system first finds the relevant fragments and passes them to the model along with the question so it can write the answer. The model does not "know" anything about your company: it reads what you give it at that moment and responds.

That difference has three practical consequences, and all of them work in your favour:

• Your documents do not go into the model. They stay in your repository. Delete a document and it stops existing for the assistant immediately.
• The answer can cite its source. Because the system knows which fragments it used, it can tell you which document and which page each statement comes from. An assistant that does not cite is not trustworthy for professional use.
• It updates itself. Change the procedure, it gets re-indexed, and the next question answers with the new version.

The expensive trap: permissions

This is the one that sinks projects, and it almost never appears in the sales proposal.

If you index "all the company's documents" into a single repository and put a conversational search on top, you have just built a machine for bypassing your folder permissions. The warehouse clerk asks the assistant about a supplier's terms and gets them. Someone asks about payroll and, if that spreadsheet was indexed, it shows up.

It is not a model failure: it is that nobody replicated the permissions. A serious document assistant has to know who is asking and search only within what that person could already open. It is the first question you should ask any provider, and the right answer is not "the model does not share information between users", but how permissions are applied to the search itself.

What to demand, specifically

Five questions that separate a serious proposal from a pretty demo:

1. Where are the documents processed and stored? Country and infrastructure provider. If the answer is vague, stop there.
2. Are my documents used to train third-party models? It has to be in the contract, not in a sentence on a website.
3. Does it respect the permissions I already have? And how: per user, per group, per folder.
4. Does it cite the source of every answer? With a link to the document and, if possible, to the page or section.
5. What happens when it does not know? An assistant that never says "I cannot find this in your documentation" is an assistant that makes things up. Ask for it explicitly and test it with a question whose answer is in no document.

What you can realistically expect

It works very well when the answer exists somewhere and the problem is finding it: the terms of a contract, how an internal procedure works, what was agreed with a client three years ago, what a regulation the company has already documented says.

It works badly, or not at all, when the answer requires counting or calculating across many documents at once ("how many contracts expire this quarter?"). That is not a documentation question, it is a database query, and it is solved differently — usually by connecting to the management system, not to the search.

And there is a precondition nobody wants to hear: if your documentation is messy, duplicated and has three versions of the same contract, the assistant will hand that mess back to you with great confidence and excellent wording. The quality of the answer cannot be better than the quality of the documents. That is the part of the project that cannot be outsourced.

If you want to see how we approach it, with what guarantees and at what price: Can AI review my company's documentation? (in Spanish).

This article is informational. Tecnea builds this type of assistant, so we have a direct commercial interest in the topic. The five questions above are written so that you can ask them of us too — and if a provider will not answer them in writing, that is the answer.