AI knowledge management: fix stale wikis with a permissioned assistant

AI knowledge management means putting an assistant on top of the documents your company already has, so staff can ask a question in plain language and get a short, cited answer drawn from current, permitted sources. It works only as well as the content underneath. If the wiki is stale, three versions of the expense policy exist and nobody owns the pages, an assistant will quote the wrong one more fluently than search ever did.

#Start with the knowledge problem, not the model

Most buying conversations begin with the tool: which assistant, which vector database. A better starting point is a list of what's going wrong today.

• Staff ask the same questions in Slack because they don't trust the wiki.
• Sales and support copy answers from old emails because the approved version is hard to find.
• A policy changed in March and half the team is still working to the old one.
• A new hire needs six people to find where anything lives.
• Security questionnaires take days because the answers sit in five different documents.

None of these is a model problem. They come from pages nobody owns or reviews. An assistant exposes them quickly, which is useful, and makes them worse if you ignore them. That's why we deal with content before technology.

#Where the knowledge sits and who may see it

Before anything is indexed, make an inventory. For each source, record what it contains, where it lives, who owns it, how often it changes and who's allowed to read it.

That last column is where the real work is. Leaving something out is a legitimate answer. Board papers, individual pay and disciplinary files are better excluded than guarded by clever rules.

#The content-ownership checklist

This is the part that makes any assistant better, including the one you already own. Run it for each source before a launch.

1. Every document or space has a named owner, a person and not a team alias, who is responsible for accuracy.
2. Each page shows a last-reviewed date, and a review cycle is set: quarterly for fast-moving policies, yearly for stable ones.
3. One document is the source of truth per topic. Superseded versions are archived or clearly marked, and the assistant is told to prefer recent content.
4. Titles say what the document is. "Final v3 (copy)" is not a title, and retrieval depends on titles and headings.
5. Tables, policies and FAQs are in a form that survives parsing: real headings, and text rather than scanned images where possible.
6. Access is set at the source (folder or space level), so the assistant can mirror it. Permissions that live in people's heads can't be enforced.
7. Sensitive categories are decided up front: included with restrictions, or excluded.
8. A feedback route exists: a thumbs-down, a "report outdated answer" button, and a person who reads them weekly.

We'd spend the first hour on items 1 to 3 before touching a prompt. Fixing them helps answer quality more than prompt tuning does, and it improves your wiki for people too.

#What the assistant should do

A good internal assistant is a cited-answer tool. It searches the documents the person is allowed to see, writes a brief answer, links the passage it came from and says plainly when it can't find anything. It doesn't guess, and it doesn't answer company-specific questions from general knowledge.

That behavior comes from a pipeline of connectors, structure-aware parsing, search that combines meaning and exact terms, re-ranking and permission filtering. The technical detail is in our guide to how a RAG knowledge assistant works. This article stays on the buyer's side: what to do to your content, who owns it and how to judge results.

Permissions deserve a line of their own. The assistant should know who is asking through your single sign-on and filter results by that identity before any text reaches the model. A shared account that can read everything is the failure to avoid, because the first person to ask about salaries will find out.

#Enterprise search, wiki, knowledge assistant

These terms get used loosely, so here is a plain comparison.

They aren't rivals. The wiki stays your content home, search stays useful for browsing, and the assistant sits on top. Microsoft 365 Copilot is permission-aware inside Microsoft 365, so if your knowledge lives there, it may cover enough. A custom assistant earns its place when answers must also draw on a help desk, a database or several document systems, or when you need a test set built from your own questions and control over where the model runs.

#Where to start

Pick one domain with a clear owner, a manageable pile of documents and frequent questions. IT and HR policy is the usual choice, and vendor security questionnaires are another. Our use-case pages show two examples: answering employee policy questions and answering vendor security questionnaires. In both, the assistant drafts and cites, and a person approves anything that leaves the company.

A realistic first month looks like this:

1. Week one. Choose the domain, list the sources and apply the ownership checklist to them.
2. Week two. Collect 100 to 300 real questions from the help channel, tickets and search logs. Have the owners write the expected answers and sources.
3. Week three. Connect the sources and score retrieval and answers against the question set. Fix the content issues the scores expose.
4. Week four. Release to a pilot group with feedback buttons on. Review unanswered questions weekly.

#How to tell whether it's working

Pick measures that point at causes.

• The share of test questions answered correctly with the right source, and the share correctly declined.
• Questions the assistant couldn't answer, which show where documentation is missing.
• Thumbs-down reasons, grouped into outdated, wrong, incomplete and unclear.
• Documents past their review date that are still being cited.
• Time to answer for the repeat questions that used to go to a colleague. Measure a before and after on a sample, and don't estimate it.

A high count of unanswered questions is a to-do list for the content owners.

Key point: An assistant can't fix a knowledge base nobody owns. Assign owners, retire old versions and set permissions at the source first, and then any assistant you choose will answer better.

#When this isn't the right project

Skip it if the knowledge isn't written down, because there's nothing to retrieve. Questions that need a calculation across many records, such as "how many contracts renew in Q3", belong to a report or a database query. And if nobody can commit to owning the content, quality will drift within months.

If you'd like help deciding where to start, read about our knowledge assistants or book a strategy call. We'll look at your sources, permissions and most common questions, and say plainly whether an assistant is worth building. The RAG glossary entry has the short definition of the method behind it.

#Frequently asked questions

#What is AI knowledge management?

It's the use of AI to make a company's written knowledge findable and usable: an assistant searches your documents, answers in plain language and cites the passages. It depends on current content with clear owners and on permissions that match each person's access.

#How is an AI knowledge base different from a wiki?

A wiki stores pages that people maintain and search. An AI knowledge base assistant reads those pages and the other sources you connect, then answers a question directly with links to where each point came from. The wiki still needs maintaining, and the assistant makes gaps in it visible.

#Is enterprise search with AI the same as a knowledge assistant?

They overlap. Enterprise search returns a ranked list of documents, and a knowledge assistant reads the top results and writes a short cited answer. Many products do both. What matters is whether it respects permissions, cites sources and declines when it can't find an answer.

#How do you keep an AI assistant from using outdated documents?

Name owners, set review dates, archive superseded versions and tell the assistant to prefer recent content. Also track which stale documents are still cited, and let users flag outdated answers. Content hygiene helps more than any model setting.

#Can an internal knowledge assistant run in our own cloud?

Yes. The index, search components and model can be deployed in your own AWS, Google Cloud or Azure account, and some organizations use hosted models under enterprise data terms instead. The choice depends on your data, your policies and your contracts.