How do you organise company knowledge before connecting AI?
A knowledge base needs owners, sources and review dates. Organise documents so that people and AI agents can find and use them.
Article language: English

In this article
If three versions of a price list circulate in your company, connecting AI search gives you faster access to three versions of the price list. Someone still has to decide which one applies. That is why organising knowledge starts with decisions about sources, owners and currency. Choose a search tool after that work.
Knowledge architecture is the way information is arranged so people can find, understand and use it. In a small company, it may mean a few well-maintained registers and short instructions. The size of the library matters less than being able to recognise a document you can rely on.
Which knowledge should you organise first?
Choose a process where the same questions keep coming back. Preparing a proposal, onboarding someone new, handling an enquiry or handing over a project can be a good place to start. Gather the material used in that process and check which sources people actually consult.
Do not move the entire drive into a new tool immediately. Start with documents needed for a few recurring decisions. For each, ask who uses it, when, and what they do on its basis. A document that nobody can connect to a task may belong in a review queue rather than on the front page of the knowledge base.
Record gaps too. Sometimes an answer exists only in one person's head. A short conversation about a specific task can provide material for an instruction. If the sequence is hard to reconstruct from an account alone, observe an expert's ordinary work (in Polish). Test the resulting description when onboarding an employee into company knowledge (in Polish): does it enable them to perform the task, rather than merely find a file?
What should a document description contain?
Use a shared description: a short record of essential details. It need not be elaborate, but it should accompany each document intended to serve as a source:
- A title describing the task or subject.
- A content owner and approver.
- A status: draft, approved or archived.
- The last review date and the next review date.
- The scope of use and links to source material.
- The access level and a way to report an error.
This description makes it easier to see what can be used in a proposal and what is still a draft. The owner need not write every update personally. They should, however, know they are responsible for resolving comments and approving the current version.
How should content be structured so people can use it?
Write around questions and actions. An instruction called “How to prepare a proposal for a new service” is easier to find than a file called “Miscellaneous decisions final 4”. Short sections with unambiguous headings let you link to the exact passage.
Separate a rule from an example and a historical decision. A rule explains how the company currently works. An example shows how it applies. A decision history explains why an option was chosen. When all three are buried in one long paragraph, yesterday's exception can easily be mistaken for today's rule.
Record limitations. If a procedure applies to one product or a particular type of agreement, put that at the beginning. Link to the source rather than copying the same price list into several documents. You can then update it in one place and check the result more easily.
How can you tell whether search helps?
Collect questions from daily work and record correct answers with their sources. Then ask people who did not build the knowledge base to perform those tasks. Observe what they find and where they start guessing. A search test does not replace checking whether the work can be completed. If only one expert can still finish a task, separately examine the company's dependence on one person (in Polish).
For AI-based search, test questions with no answer, conflicts between versions and material unavailable to a given user. An answer should let the person open the source. When evidence is missing, it should clearly signal that the information needs confirmation.
Keep the test questions after launch. They will be useful when documents move, permissions change or the tool is replaced. Compare results with the agreed answer and record errors so you can identify the cause: a missing document, outdated content, incorrect access or the search method.
When does knowledge start to age?
Often after a change to a product, price, supplier or responsible person. Tie reviews to events in the company. A quarterly calendar alone may not be enough if a new offer launches tomorrow.
Keep archived material clearly labelled and limit its accidental use in current work. If an old document remains searchable, it should point to its successor. Important changes also need a record: what was corrected, who approved it and when it takes effect.
NIST highlights the need to document roles, monitor and manage risk throughout an AI system's lifecycle. That is useful context when assigning responsibility for the knowledge base after launch. See the NIST AI RMF. Organising documents alone does not constitute a compliance assessment of the whole system.

How do you resolve a conflict between two plausible documents?
Prepare a small discrepancy record. Enter the user's question, both answers, the source of each and the person authorised to decide. Establish scope first: a price list for a new client can differ from terms in an older agreement. A newer file date alone does not prove that the document replaces every earlier commitment.
Consider a working example: a sales instruction gives one level of support and a proposal template gives another. The service owner checks which orders each applies to, updates the instruction, marks the exception and adds the condition for using the older template. The person preparing a proposal gets one starting point and a link to the details. They do not have to choose between attachments on their own.
Keep the reason in a project decision register. If the conflict concerns data sent to a model, add it to the AI data-flow register too. Correcting the document can then prompt a check of places where its copy may already have been used. For preparation, see the guide to preparing a project decision register (in Polish) and the guide to preparing an AI data-flow register (in Polish).
How do you test the knowledge base across permission levels?
Prepare pairs of tasks for several roles. A new sales employee looks for proposal rules. The service owner checks internal pricing conditions. An external collaborator needs only instructions for handing over materials. Each role should receive an answer within its granted access, together with the appropriate source.
Run the test using agreed test accounts and material approved for that purpose. Check search results, document previews and links opened from answers. Record what was visible and from which account. A promise of access control needs a test in the actual tool; the name of a feature in a supplier's offer does not prove it works.
Prepare the list of sources, accounts and restrictions using the index of material for an AI audit (in Polish). Here the boundary between roles matters more than the number of documents. When a user reports that an answer is missing, first check whether the material should be available to them.
How do you distinguish a content error from a search error?
For each unsuccessful question, check four things in order: does the answer exist, has it been approved, does the person have access, and did the tool find the right passage? Only then assess how the model formed the sentence. Otherwise the team may revise the model instruction when the cause is a missing appendix.
In a test sheet, retain the question, expected source, received answer, error type and proposed correction. Include questions for which the company does not yet know the answer. A correct result may identify the owner of the issue or ask for clarification. An invented resolution is an error even when it sounds fluent.
Repeat the same test after a change. Add new questions separately so the comparison does not disappear beneath a new version of the set. The AI pilot evaluation criteria (in Polish) develop how to record results and the conditions for the next decision. When handing over the solution, also use the code and documentation acceptance checklist (in Polish) so the company retains the test set.
When should you narrow the knowledge base rather than add files?
If most problems come from one process without an owner, stop expanding it. First approve the rule and describe the exceptions. Adding another department will not remove the dependency. It may instead increase the number of people receiving an unclear answer.
A useful boundary for a stage is whether someone who did not create the documents can perform the agreed set of tasks. Check whether they can find the source, recognise its status and report an error. Only then add the next process with its own owner. This sequence reflects the idea of roles and reviews described by NIST; it is not a certification claim. The source documentation was checked on 27 October 2026.
What can you do in the first week?
Choose one process, name an owner and organise the most frequently used sources. Add document descriptions, resolve uncertainty about the applicable version and test with real questions from the team. The exercise will show whether the problem is missing information, its structure or access.
Choose the next step from the result. It may be completing an instruction, improving navigation or introducing search. If an agent will act on this content, read about boundaries for AI agents (in Polish). The Audit service covers an assessment of AI tools and information flows.
Aeneia Audit: Start with one task, the sources used by the team and its AI tools. An agreed assessment can reveal gaps in information flow and priorities for organising it. Discuss the knowledge used by AI.
If documents lose the reason behind earlier decisions, add a project decision register. It preserves context, alternatives and conditions for change without rewriting the entire knowledge base. For preparation, see the guide to preparing a project decision register (in Polish).
Organising also includes archiving outdated company knowledge (in Polish). Decide what leaves current search, how to identify a successor and who confirms the change. An old instruction may remain useful evidence of history, but it should not accidentally present itself as the current rule.












