Knowledge Centre: Owners, Upkeep and AI With Sources

Many companies have a wiki nobody trusts. Pages are outdated, duplicated or have no contact person. That is rarely a problem with the tool. What is missing is structure and responsibility.

The article on AI search across company knowledge was about making existing knowledge findable. This one is about the next step: a knowledge centre that stays current and that an AI can rely on.

The starting point

According to a 2023 Gartner survey, 47 % of digital workersU struggle to find the information they need to do their jobs. On average they use 11 applicationsU, up from 6 in 2019. Gartner surveyed 4,861 workers in the USA, the UK, India and China, not in Germany.

In its State of Teams 2025 survey, Atlassian puts the time teams spend searching for answers at 25 % of working timeU. Atlassian itself sells software for this problem. The research institute APQC measured 2.8 hours a weekU spent searching for or asking for information in 2021.

The framework: four page types

The Diátaxis framework distinguishes four kinds of documentation. Every page has exactly one type, and every type has a template.

Page type	Example	Review cycle	Owner
Learning	Your first week in sales	12 months	Team lead
How-to	Submitting travel expenses	6 months	Department
Reference	Price list, policy, contacts	with every change, at least every 6 months	Department
Background	Why we price the way we do	12 months	Division head

On top of that, every page has five mandatory fields: owner as a named person, status, review date, audience with permissions, and source. No page goes live without them.

Roles

Owner: One person per page or topic area, not a team. They confirm the page by the review date or revise it.
Knowledge lead: A small role for standards, templates and metrics, not a whole team.
Everyone else: They capture knowledge where it arises, that is, while solving a problem.

This principle comes from the KCS method, renamed Knowledge-Centered Success in April 2026. Knowledge is created in the flow of work, and every use is also a review. Pages move through states such as "work in progress", "not validated" and "validated". Members of the consortium report 50 to 60 % shorter resolution timesH, though without a disclosed methodology.

If you are looking for a formal framework, the ISO 30401 standard for knowledge management systems provides one. A revised second edition is in draft and has not yet been published.

The workflow

WorkflowAI stepPersonSystem and dataCheck

01StartA question comes up, for example in support or the team chat
02AIAnswer only from verified pages, with source and review date
03CheckAnswer found and correct?
no: report the gap, a new page is created in the flow of work
04PersonOwner reviews the new or changed page and sets the status to verified
05SystemSet the review date, reminder four weeks before it expires
06DataMetrics: searches without results, wrong answers, pages past their review date
07ResultKnowledge stays current and findable

The AI layer: answers only with sources

The AI answers only from verified pages and names the source for every statement. If it finds nothing, it says so, and the gap goes to the responsible person. That way, every unanswered question becomes an assignment for a new page.

Prompt

Answer the question using only information from the provided knowledge centre pages. For every statement, name the page with its title and review date. If a page's review date has passed, point this out. If you find no matching page, reply: "There is no verified page on this yet." and name the responsible person if the pages mention one. Do not add anything from general knowledge.

Two risks remain. The first is permissions. According to Microsoft, Microsoft 365 Copilot only shows content a person is already allowed to see. SharePoint pages that are shared too widely therefore suddenly appear in answers. Microsoft offers a dedicated guide against oversharing and reports on overly shared sites. According to the vendors, Atlassian Rovo, Notion AI and Glean also respect existing permissions.

The second risk is wrong answers despite sources. A Stanford University study found hallucinations in 17 to 33 % of answersS from legal research tools that cite sources. It covered legal questions and the 2024 versions. So display sources prominently and offer a "wrong answer" feedback button.

Maintenance: review dates instead of clean-up sprints

Review date: Every page has one. Four weeks beforehand, the system reminds the owner.
Expired pages: They are visibly marked as unverified and no longer feed into AI answers.
Archive rather than delete: Except for pages with personal data. The GDPR's storage limitation applies to them, and they need a deletion deadline.
Metrics: Searches without results, answers flagged as wrong, pages past their review date and the most-used pages.

Law and the works council

Co-determination: If the system logs usage, Section 87(1) No. 6 of the German Works Constitution Act applies. A works agreement is generally required.
Trade secrets: Germany's Trade Secrets Act only protects information if appropriate confidentiality measures are in place. Clean access rights in the knowledge centre are therefore also legal protection.

Limits and risks

Tool before structure: Starting with the tool builds the next wiki nobody trusts.
Oversharing: The AI makes content that is shared too widely visible.
Where the numbers come from: The figures come from surveys without a German sample or from vendors with their own product. The KCS figures are self-reported.

How to start

Choose an area with many recurring questions, such as HR or IT support.
Collect the 30 most frequent questions from tickets and team chat.
Write or revise the pages according to the four types, each with an owner and review date.
Enable AI answers for this area only, with sources and a feedback button.
After eight weeks, measure searches without results and wrong answers. Only then move on to the next area.

A Knowledge Centre That Stays Maintained: Structure, Roles and an AI That Answers With Sources