Everything so far assumed you paste text into a box. The bigger shift is pointing AI at material *you* own — the policy manual, the last three years of board minutes, a folder of supplier contracts, the 200-page standard your industry runs on.
This is where AI stops being a clever writing toy and becomes useful for the specific job you actually have. It is also where a new failure mode arrives.
Why your documents change everything
Remember the sorting rule from lesson one: the model is strong when truth lives in the text you gave it. Uploading your documents *puts the truth in the text.* A question the model would have guessed at — "how many days of bereavement leave do we give?" — becomes a question it can answer by reading.
Three ways this shows up in tools you already have:
- Attach a file to a chat. Simplest. Works for one to a few documents. The model reads the whole thing.
- A project or knowledge space. Most assistants now let you keep a set of documents attached across many conversations, so you stop re-uploading the HR manual every morning.
- Search over a large collection (often called retrieval or RAG). For hundreds or thousands of documents, the tool finds the relevant passages first, then answers from those. Important consequence: it only sees what the search found. If the search misses, the answer is built on an incomplete picture and will not tell you so.
The failure mode: confident blending
When you supply documents, the model does not cleanly separate "what your document says" from "what I generally know about this topic." Ask about notice periods with your handbook attached, and you can get an answer that is 80% your handbook and 20% standard employment practice, seamlessly merged, in one paragraph, with no seam visible.
This is worse than an obvious error because it is mostly right. The fix is structural, not hopeful:
Answer only from the attached documents. For every claim, give the document name and the quoted sentence it comes from. If the documents do not answer the question, say "not covered in these documents" — do not fill the gap from general knowledge.
Then spot-check two or three citations. Citations can be wrong. A quoted sentence that does not exist in the file is the single fastest way to catch a bad answer, and it takes one search to find out.
What this is genuinely good for
- Finding, not reading. "Which of these 40 supplier contracts have a price-escalation clause?" You then read those four.
- Consistency checks. "Does the new procedure contradict anything in the existing manual?" Humans are poor at this across long documents. Machines are decent.
- Answering the same question for the tenth time. Shopkeepers with a returns policy, teachers with an exam board specification, civil servants with a procurement rulebook — the question comes in weekly and the answer is in a document nobody wants to reopen.
- Onboarding yourself. New to a case, a client, a ward, a portfolio? "Here are the last eight months of notes. What do I need to know before Monday?"
Before you upload anything
Stop. Whose documents are these, and what did you promise about them? A client contract, a patient file, a personnel record, anything under an NDA — the answer may be that you cannot upload it at all, regardless of how useful it would be. That is lesson six, and it is the one that ends careers. Read it before you upload your first real file.
Before you move on