--- title: "How to Prepare Company Data for AI" description: "Most failed AI projects fail on data, not on the model. What \"preparing the data\" means in practice and how to tell whether a company is ready at all." canonical: "https://interfase.sk/en/news/how-to-prepare-company-data-for-ai" lang: "en" last_updated: "2026-10-28" --- # How to Prepare Company Data for AI > Most failed AI projects fail on data, not on the model. What "preparing the data" means in practice and how to tell whether a company is ready at all. Published: 2026-10-28 - The quality of AI answers follows the quality of the inputs; a model will not resolve a contradiction between two versions of the same document. - The first step is not cleanup but an inventory — where the data is, who owns it and which of it is binding. - Permissions have to be part of the data, or the agent will expose to everyone everything it can reach itself. - Do not prepare the whole archive at once — pick one area, finish it properly, then expand. "We want AI to answer questions from our internal documents." It sounds like a two-week brief. Then it turns out the documents live in three systems, four versions of the price list disagree, and nobody in the company can say which one applies. The model will not resolve that. Given contradictory material, it will answer confidently based on whichever came to hand first. ## Start with an inventory, not with cleanup The first step is not tidying. It is finding out what you actually have. For every data source write down four things: **where it is** (system, drive, mailbox), **who owns it** (who decides what is correct in it), **how often it changes** and **whether it is binding** — that is, whether the company actually acts on it, or whether it is only working material. That table tends to be the first moment the real scope of the project becomes visible. It often shows that half the sources are duplicates and a third are out of date. ## One version of the truth per area This is the single most important decision in the whole preparation. For every topic — price list, product documentation, internal policies, contract terms — there has to be **one source that applies**. The rest are either retired or unambiguously marked as superseded. Without that, nothing reliable can be built. Not because AI is weak, but because "how much does X cost" has no single answer even for a person. ## What cleanup means in practice Once it is clear which sources apply, the work moves to their content: - **Remove what is out of date.** A 2019 document describing a discontinued process is worse than no document. - **Add a validity date.** So current can be told from historical by machine as well as by eye. - **Break up large documents.** A 200-page policy cannot be used meaningfully as one block; splitting it into chapters with their own headings improves accuracy considerably. - **Deal with scans.** A PDF that is an image is empty as far as AI is concerned. Either convert it to text or leave it out. - **Unify naming.** If one product is named slightly differently in three systems, you need a mapping. ## Permissions belong to the data This gets underestimated and it is the riskiest part. If the agent sees everything a system administrator sees, then **every user sees the same through it** — including salaries, personnel files and deals in progress. Permissions are therefore handled at the data layer, not by instructing the model. Every document carries information about who may see it, and filtering happens before anything reaches the model. We describe the related principles in the piece on [AI agents and GDPR](/blog/ai-agenti-a-gdpr-na-co-si-dat-pozor-pri-spracovani-osobnych-udajov). ## Metadata decides accuracy Text alone is not enough. Attaching **title, department, date, document type and language** to each source lets the search narrow down before it looks inside the content at all — and the answers get markedly more accurate. In a multilingual company, language is mandatory; we explain why in the piece on [multilingual support in a business application](/blog/viacjazycnost-a-lokalizacia-firemnej-aplikacie). ## Structured data follows different rules With data from databases and ERP it is not about cleaning text but about **access**. The question "what was last month's revenue" should not be answered by handing the model raw tables. It should get a prepared, verified query or report that returns the same figure as the company's own statements. Otherwise you get the worst possible outcome: two numbers for the same thing and no certainty which is right. ## Do not prepare everything at once The most common mistake is trying to tidy the entire company archive before the first deployment. It takes months and nobody sees a result. The opposite works: pick **one area with the most frequent questions**, finish it properly, deploy it and measure. The first area teaches the company what order looks like — and the second then goes three times faster. ## How to tell the data is ready A simple test: take twenty real questions people in the company ask, and try to answer them **from the prepared material alone**. If someone unfamiliar with the area can manage it, so can an agent. If that person has to phone somebody, the data is not ready. If you are working on this and want to go through what is realistic in your case, [get in touch](/kontakt). More on [AI and automation](/riesenia/ai) and examples in our [case studies](/referencie). --- Web version: https://interfase.sk/en/news/how-to-prepare-company-data-for-ai llms.txt: https://interfase.sk/llms.txt