"My business is very specific": how to make the AI answer from your own documents

The short answer

You load your business documentation and make it answer from there. An AI that answers from my documents does not generate a plausible reply: it retrieves the paragraph that answers the question and answers with that.

It is the difference between an AI that knows your industry and one that knows your business. Anyone can have the first. You build the second, and it takes an afternoon.

Why "my business is very specific" is true

It almost always is, and it is not an excuse to avoid automating.

Your returns policy has an exception you invented. Your delivery times depend on three zones you drew. Your price list has a volume discount written down nowhere except a PDF and the heads of two people.

None of that is on the internet. A model trained on the internet cannot know it, and if asked, it answers anyway. Stanford's AI Index 2026 reports a benchmark across 26 frontier models with error rates from 22% to 94%: without a source of yours, the model guesses.

What to load, and in what order

Order matters because the first document already kills half your repeat questions.

Order	Document	Questions it silences
1	Current price list	"How much is it?", "is there a volume discount?"
2	Shipping policy and lead times	"Do you deliver to my area?", "when do I get it?"
3	Returns, exchanges and warranty	"Can I exchange it?", "until when?"
4	Real frequently asked questions	Everything your team has answered a thousand times
5	Product sheets and manuals	The technical ones, the ones that escalate today

The fourth costs nothing and almost nobody does it: open your WhatsApp, read the last hundred conversations, and write down the ten questions that repeat most with their answers. That document pays the best.

How it searches, and why it is not a keyword search

It finds the passage, not the file.

If a customer asks "will this survive the rain?" and your spec sheet says "IPX4 rated", a keyword search finds nothing. A meaning-based search does, because it understands they are the same thing.

That is why you can upload the forty-page manual nobody will read end to end. The AI does not hand back the manual. It hands back the paragraph.

What happens when you change a price

You change the document. Nothing gets retrained.

This is the part that confuses most, because "training" is the word in circulation. The documentation does not live inside the model, it lives beside it. You upload the new list and the next answer already carries the new price.

If the figure lives in your management system and changes constantly, there is a better route than a document: connect the catalog, so price and stock are read from there at the moment of answering.

With one caveat worth saying up front rather than later: today the system you connect from the panel is Dux Software. Others are added one at a time. If you run something else, for now the route is the updated document.

How you verify it quotes instead of improvising

Three tests, before it talks to a customer.

1. Ask it something that is in the document. It has to answer with what the document says, not a near approximation.
2. Ask it something you never loaded. It has to say it does not have it.
3. Assert something false: "I was told you ship free". It has to correct you.

If all three pass, you know where every answer comes from.

What it does not fix

Loading documents does not make the AI infallible, and that is worth saying.

There are still enquiries that need a person: a complaint, a commercial exception, a case your documentation never anticipated.

A building management company we serve receives around 312 conversations a month, and the overwhelming majority fall into four repeating subjects: leaks, noise, lifts and maintenance. That handful is exactly what fits in a document and stops occupying your team.

What does not fit is the exception. Documents raise how much the AI resolves on its own. They do not make it infallible.

Key takeaways

• An AI without a source of its own guesses, and measured error rates across top models run from 22% to 94%.
• Documentation is loaded, not trained: you change the file and the answer changes.
• Start with the price list, then shipping, returns, and the ten questions your team repeats most.
• Search works by meaning, so the long manual nobody reads end to end is still worth uploading.
• Verify with three tests: that it quotes, that it admits what it does not know, and that it pushes back on a false claim.