Why AI on top of chaos does not work
Companies often say: we will deploy AI and it will somehow figure things out. But AI finds its way around documents no better than a new employee — in some ways worse, because it does not know the 2022 price list is no longer valid. Hand it three versions of one policy and it answers from whichever it found first, and it sounds convincing.
You know the typical signs of chaos: outdated documents nobody deleted; duplicates in customer data; files scattered across emails, personal drives, and USB sticks; and most often, knowledge that exists only in people's heads. “Only Jana knows that” is a sentence AI cannot fix, because Jana never wrote it down.
The conclusion is unpleasant but useful: you cannot buy better AI answers with a better model. They are capped by the quality of what the AI draws from. Order in your source material is not a formality to tick off before AI — it is a substantial part of the whole job.
First, find out where knowledge really lives
The first step needs no software. Take a sheet of paper or a shared spreadsheet and write down where the important information lives in your company: contracts, price lists, procedures, answers to common customer questions, technical guides, internal rules. Next to each, note the place — a shared drive, a mailbox, a system, a personal computer, a specific person's head.
Ask people on the team a simple question: “When you need to find this, where do you look and who do you ask?” The answers are often surprising. You will find the price list exists in four versions, that the complaint procedure lives only in an email from last year, and that if one person left, the company would lose the way a key customer is invoiced.
The result is not a perfect catalogue. It is a rough map that shows where the duplicates and gaps are, and where the company is vulnerable even if it never deploys AI.
Pick one area and remove the outdated
The temptation to tidy everything at once is strong. Do not do it — after three weeks you run out of steam and the company drifts back to the old ways. Choose one area where the pain is greatest and where order can be maintained: customer support material, internal procedures, or price lists and quotes, for instance.
Then comes dull but useful work: go through the documents and ask one question about each — is this still valid? Whatever is not, move to an archive, not the bin, so you have peace of mind. Of several versions, keep one and name it clearly, for example with a valid-from date. Merge duplicates. A file naming rule fits on half a page.
This one area also serves as a trial run. You will see how long the tidying took, what turned out to be hard, and whether it can be kept up. Only then does it make sense to move on to the next.
Name an owner and write down the unwritten
Order falls apart without a person behind it. So for each area — not each file — name an owner: one person responsible for keeping the content current. They do not have to write everything themselves. They need to know that when a price list or procedure changes, it has to be updated, and that twice a year they will check whether everything still holds.
The other half is writing down what is currently only in heads. It does not have to be a book. A short conversation or recording helps: have an experienced colleague say aloud how they handle a typical situation, and turn it into a step-by-step procedure. The first version will be rough, and that is fine — the first person who works from it will correct it.
This is how tacit knowledge becomes a company asset. A new hire gets up to speed faster, a key person's holiday is not a crisis, and if you later decide on AI, it will have something to draw from.
Who may see what, and what must not leave
Before you give documents to any AI tool, decide two things. The first is access rights. Not everyone on the team should see salaries, supplier contracts, or correspondence with a lawyer. If AI reads everything, it can easily tell someone something they had no right to know. You therefore need a clear split of who has access to which area — and that can be done without AI.
The second is what must not leave the company. Customers' personal data, trade secrets, individually negotiated prices, health or financial information. Sort documents into at least three groups: open, internal, and those that do not belong in an external AI service at all, or only after editing. We covered what leaves a company when it deploys AI in a separate article; here it is enough to say this decision must be made before the first document is uploaded.
This is not bureaucracy for its own sake. It is the same question you should ask when emailing a document to a stranger.
What you gain even without AI — and where it does not pay off
The good news: every one of these steps pays off even if you never deploy AI. People find what they are looking for faster. Fewer jobs get done from an old price list. Onboarding new colleagues is shorter and the company depends less on a single person. And if an AI that answers from your documents does arrive — we wrote about that in the RAG article — it gets a prepared foundation, not a mess.
It is also true that preparation should not drag on forever. If the task is narrow and clearly bounded, with a few pages of source material, you do not need months of cleaning; going through those pages is enough. And if you are not thinking about AI at all, do not invent order just because it is fashionable — create it where it actually slows you down.
If you would like advice on mapping and tidying your material, we help with that at Lucorp. But the first three steps — a map, one area, an owner — you can manage yourself, with a sheet of paper and a bit of patience.