Preparing documents for an AI assistant: what to fix first

An AI assistant knows nothing about your company that you haven't given it. If three versions of the same procedure sit together in the documents folder, the assistant will pick one: not the most recent, but the one best written for the question it received. A customer will hear last year's terms, and the answer will sound just as confident as any other.

That's why the real work starts before the system: putting in order what exists, which version is valid, who can see it, and what happens when there's no answer. This work is useful regardless, even if the assistant is never built: it's the same stuff that slows down new employees.

Inventory: what exists, who's responsible for it, which version is valid

The first step is a list. Not of folders, but of the documents that contain answers: manuals, procedures, terms of sale, technical sheets, answers already given by support. For each one you need four pieces of information.

The most useful rows are the last two. A document no one is responsible for isn't a source: it's archive material. And a document with no date can't be judged, so it can't be used to answer a customer.

Public, internal, confidential: the separation comes first

Classification isn't a bureaucratic formality: it's what stops a customer-facing assistant from quoting the reseller price list. Three levels are almost always enough.

• Public: something you can say to whoever writes in. If you wouldn't say a piece of content to an unknown customer on the phone, it isn't public.
• Internal: useful to employees, not to customers. Margins, special terms, procedures that reveal internal organization.
• Confidential: personal data, individual contracts, information about other customers. As a rule, it doesn't go into the assistant's sources, and if it does, it needs its own access controls.

The practical rule is that the separation lives in the permissions, not in the instructions. An assistant told not to reveal the price list, but that still has the price list among its available sources, will use it sooner or later: all it takes is a question phrased at an angle. A confidential document must be unreachable, not merely discouraged.

Documents written to be read in pieces

An assistant doesn't read a manual from start to finish: it retrieves the passages that seem to answer the question. A document meant for continuous reading behaves badly, because its pieces don't stand on their own.

• Real, specific headings. A section called "General" doesn't say what it's about; "Warranty on replaced parts" does.
• Every section must make sense taken on its own. Implicit references — "as mentioned above," "in this case" — become meaningless the moment the piece is extracted.
• The information that matters shouldn't sit inside an image. A photographed price list, a scanned diagram, or a document made of scanned pages are, as far as the system is concerned, blank pages.
• Tables should stay tables, with headers repeated if they're long.
• Terms and conditions need to be written out in full where they're needed: if the rule on returns depends on an attachment that isn't among the sources, the answer will be incomplete and no one will notice.
• A short, accurate document beats a long, generic one: length doesn't help retrieval, precision does.

What should happen when the source isn't there

This is the decision that separates a reliable system from a dangerous one, and it should be made by you, not left to the default settings. The options are few: answer that the information isn't available, offer to connect with a person, or open an internal request.

Who updates it, and how often

A knowledge base decays just by sitting still. Terms change, a product is discontinued, a procedure gets replaced, and the assistant keeps answering the same way as before, with perfect confidence.

1. Every document has an owner with a first and last name, not a department.
2. Every document has a review date: when it needs to be looked at again even if no one flags anything.
3. There's exactly one place where it gets updated. If the document lives in two places, one of them will be out of date, and you won't know which.
4. When a document is replaced, the previous one leaves the sources at the same moment. Renaming it isn't enough: as long as it's reachable, it will be used.
5. The questions the assistant couldn't answer get reviewed at regular intervals: they're the list of missing documents, written by your customers.

If this work isn't sustainable — no one can take it on, no one can say which version is valid — that's a good reason to postpone the project or narrow its scope to a single area, the one where the documents are in order. An assistant built on sources no one maintains makes the situation worse: it gives wrong answers faster than before.

What this guide doesn't cover

Here you prepare the knowledge: what goes into the sources, in what form, and with what permissions. How you then verify the assistant's behavior — the tests to run before putting it online, checking the answers, handling errors — is a separate, later job, covered elsewhere. One distinction is worth remembering: making documents available to a system that retrieves them isn't the same thing as training a model on your data, and the two have different implications.