Keywordsai for knowledge managementBlogai for workai at workworkplace aiai for professionalshow to use ai at workai productivity
Related searchesai for knowledge managementchatgpt company wikirag for internal docsai knowledge base at workowned corpus for aihow to use ai for knowledge management
A chatbot on a junk drive will sound certain
Knowledge management with AI fails in a predictable way. The organisation dumps a share drive into an index, puts a chat box on the intranet, and announces that people can now ask the company anything. The box answers with confidence. The page it found was last edited by someone who left two years ago. The current rule lives in a slide that was never indexed. Staff follow the chat because it is faster than asking a person. The person still gets the incident. AI for knowledge management only works if the corpus is current, owned, and searchable. The chatbot is the last layer. The first layers are the ones you already neglected.
A company wiki is not a model. It is a set of pages with authors, review dates, and a place that counts as current. ChatGPT can help you draft a page from notes. RAG can help you retrieve the page. Neither creates ownership. If you cannot name who weeds the library, you are not doing knowledge management. You are doing a demo of search. Demos impress steering groups. They harm new joiners who treat the first fluent paragraph as the rule. Onboarding is where this shows first. A wrong access path taught on day two becomes muscle memory. Muscle memory is expensive to unwind.
This essay is the operations view. The RAG essay explains retrieval. This one explains why retrieval is only as honest as the shelf. You will see the three tests, how to use AI on a corpus you trust, and how ownership stays human. You will not see a promise that the model will keep the wiki tidy. Tidy is a meeting with an owner and a deletion right. Deletion is still the rarest skill in knowledge work. AI does not add it. It only makes the undeleted page more charismatic.
If a page has no owner and no review date, it does not belong in the index. Indexing it teaches the model to speak with a dead author's confidence.
Current, owned, searchable: the three tests
Current means a review date and a rule that pages past that date fall out of the index or carry a warning. Owned means a role who can change and retire the page. Searchable means titles, headings, and access that retrieval can actually use, including permissions so people do not see files they should not. Fail current and you retrieve history as policy. Fail owned and nobody fixes the miss. Fail searchable and the right page exists while the chat cites a near-match. Near-matches are how RAG goes wrong while looking technical. Technical success with the wrong file is still an operational failure.
Do not index everything because storage is cheap. Storage is cheap. Attention is not. A junk drive in the index is a certainty engine for rumour. Start with the libraries you would show an auditor: policies, product rules, live procedures. Add after those pass the three tests. ChatGPT as a front end to a neglected wiki will not discipline authors. It will hide the neglect. Hiding neglect is the opposite of knowledge management. Knowledge management is making the current page easier to find than the charming old one. If the old page still wins, you have not finished the weed.
| Test | Pass | Fail |
|---|---|---|
| Current | Review date, stale pages out or flagged | 2019 deck ranked as policy |
| Owned | Named role who can retire the page | Orphan files that still retrieve |
| Searchable | Clear titles and lawful permissions | Right page hidden, near-match cited |
| Cited | Answer with a link a human can open | Fluent paragraph, no page |
Use AI on a corpus you actually trust
Name the libraries in scope. Assign owners. Set review dates. Only then enable retrieval. Instruct the system to cite and to say not found. Use ChatGPT or an approved writer to draft new pages from owner notes, then have the owner publish in the wiki, not in the chat. Onboarding packs should point to the live pages, not to a generated handbook that will drift. Measure misses: questions that retrieved the wrong file or found nothing. Misses are the weed list. The weed list is the work. A dashboard of queries is not the work unless someone acts on the misses.
Keep a human editor for tone and conflict. When two pages disagree, the model will smooth them. Smoothing is how a dispute becomes a false standard. The owner of the corpus resolves the dispute in the pages, then retrieval will follow. Do not ask the chat to pick a winner. Picking a winner is governance. Governance needs a name. Put the name on the library, not on the vendor contract. If two owners will not meet, do not index either page until they do. A frozen conflict is better than a fluent fake resolution.
Role: You are a knowledge editor working only from the pages I retrieve or paste.
Task: Answer the question or draft a wiki update from those pages.
Context: Library name, owner, review dates, and the audience such as a new joiner.
Constraints:
- If pages conflict, list the conflict. Do not smooth it into one rule.
- If the answer is not in the pages, write not found.
- Cite the page title and date for every claim.
- Do not invent owners, systems, or review dates.
Output: Answer or draft page, citations, conflicts, and missing fields.
Quality checks: List ways a new joiner could follow the wrong version if we publish this.If conflicts appear, stop generating and call the owners. The model has done its job by refusing to pick. Your job is the meeting. Publishing a smoothed page is how knowledge management creates a third version that nobody wrote and everybody will follow. Put the conflict on the weed list with a date. If the meeting does not happen, keep not found in the chat rather than a blended rule that will be treated as policy.
Ownership theatre and stale pages that still rank
Ownership theatre is a team inbox in the footer and nobody who weeds. Stale pages keep ranking because they are longer or better titled than the current note. The chat prefers them. People prefer the chat. The current note dies. Your mitigation is ranking rules that prefer review date, and a monthly weed with deletion rights. If legal will not allow deletion, archive with a banner that says not current. Banners are ugly. Ugliness is a feature when the alternative is a confident wrong answer. Prefer the ugly banner to the charming dead page.
A second risk is permission leakage. Retrieval that ignores access control will teach the model to quote a file the user could not open in the drive. That is a data incident dressed as helpfulness. Check that the RAG path respects the same permissions as the source system. A third risk is generating a parallel handbook for onboarding that is not the wiki. Parallel handbooks drift by week two. Point people at the live page. Use AI to help them find it, not to replace it with a souvenir PDF.
| Mistake | What it looks like | What to do instead |
|---|---|---|
| Index the junk drive | Everything searchable, nothing trusted | Auditor libraries first, then expand |
| Smooth conflicts | One invented standard | List the conflict, owners decide |
| Chat as wiki | Pages that live only in answers | Publish in the owned system |
| Parallel handbook | Onboarding PDF that drifts | Links to live pages only |
A chatbot that cannot say not found will never reveal that your knowledge is rotten. It will only sound sure. Sure is not the same as current.
Related reading on StudyGrid
Read next: What Is RAG for Business Teams AI for Onboarding Advanced AI. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.
What to do this week
Pick one library you would show an auditor. Name owners and review dates. Take it out of the index if it fails either test. Put it back when current, owned, and searchable all pass. Ask five real questions from this week's work and open every citation. Log not found and wrong-file hits for the owner. That is AI for knowledge management, starting with the shelf rather than the chat box.