AI knowledge management is a cycle, not a tool: a source is selected, legally secured, distilled into patterns, approved by an expert, returned to daily work and reviewed on a fixed rhythm. The AI takes over the searching, sorting and comparing; the decision about which knowledge is valid stays with a person who has a name. It rarely fails because a chatbot is missing. It fails because nobody decides which knowledge is valid, who reviews it and when it expires. A large amount of experience lives in conversations, in customer questions and in explanations from experienced staff that never reach a manual. The path from conversation into the business has six stages and four roles.
In my business, knowledge management starts with the conversations I have anyway. The reviewed intermediate product is the pattern register from call analysis: recurring questions, each with my clearest answer so far. This article is not about the technology behind it but about the operating framework: who reviews, what is valid, when it expires and how it gets back into daily work. Which role handles which task in my business is under AI employees.
Explicit knowledge and conversation knowledge
For practical purposes I divide a company's knowledge into two kinds. This is my working distinction, not a textbook definition.
Explicit knowledge already exists in processes, proposals, handbooks and product data. It is relatively easy to collect.
Conversation knowledge appears in use. Why is an exception made? Which explanation does a customer understand immediately? Where does almost everyone ask a follow-up? This knowledge is valuable but highly contextual, and it is in no document because nobody ever wrote it down. Fraunhofer IAO describes the problem for companies this way: through demographic change, staff turnover and project changes they increasingly lose critical experiential knowledge, in particular implicit skills, networks and troubleshooting know-how that until now could hardly be documented. A call transcript does not make this knowledge usable yet; only the reviewed register separates what recurs from conversation noise.
Capture expert knowledge before it leaves the business
The question “How do I document expert knowledge?” has an uncomfortable answer: not by asking the expert to write it down. Someone who has been able to do something for fifteen years usually no longer knows which part of it needs explaining; what she knows shows when she explains it to someone. That is why the first decision is which type of conversation you analyse:
For the handover of an experienced colleague, the interview is the most productive form: one hour in which a new colleague asks and she explains, recorded with her consent, then analysed like a customer conversation. What has to be settled legally before recording is in the legal rules for transcribing conversations.
AI knowledge management in six stages
1. Select the source
Start with one clear conversation type, such as product support calls or sales calls for a single service. Importing every recording creates more risk than value.
2. Establish rights and protection
Consent, legal basis, purpose, access and retention come before recording. Local transcription can shorten the data path. Replacing names with roles afterwards is pseudonymisation at first; the entries remain personal data while attribution is possible. I am a developer, not a lawyer, and this is not legal advice; when in doubt, have your specific case checked.
3. Distil patterns
Capture questions, objections and answers with source locations: verbatim quote, trigger, call ID, minute. Merge similar statements only when the same answer fits. Keep uncertain matches flagged. This is where the AI works, and where it is checked: one sample against the transcript per batch.
4. Approve the knowledge
Every entry needs a human owner. AI may suggest, compare and flag contradictions. Expert review creates validity: the owner reads answer and source location and decides whether it holds, holds with limits, or stays “open.”
5. Return knowledge to work
Knowledge becomes valuable when it appears in a workflow: onboarding, support, FAQs from customer conversations, sales or internal search. Every entry therefore gets a field “used in.” If it is still empty after a quarter, the entry is superfluous or nobody found it.
6. Review use and decay
Record the last review, next review and dependent assets. Frequently corrected answers and failed searches become the next maintenance queue. In my business this runs quarterly.
Roles instead of an ownerless shared folder
Each knowledge area needs at least four responsibilities:
One person can hold several roles in a small company, but the responsibilities must remain visible. “The AI handles it” is not ownership. An AI employee can carry a large part of the knowledge steward role, because that is the monotonous one; source owner and expert owner remain people with names.
One entry goes through the cycle
The following example is constructed; every person, number and statement is fictional. An installation business with twelve staff wants to secure the knowledge of its master craftswoman before she retires in eighteen months. The business chooses handover interviews: one hour a week, a new journeyman asks, the master explains, recorded with her consent.
The most important yield is the two open entries. “I do that by feel” is precisely the knowledge that would leave the business, and now it sits in the register as a question, with a date.
Which technical level fits?
No level is automatically more mature than another. The smallest system that reliably answers the real question is often the best choice. The demo business above gets by with the register as long as sixteen entries sit in an onboarding folder.
Quality rules for AI knowledge management
- Answers need sources. Without evidence, a statement remains a draft.
- Contradictions stay visible. AI must not average them into a convenient answer. If the master says something different in interview 3 than in interview 8, both statements sit in the entry and the expert owner decides.
- Validity has a date. Prices, law, products and processes change.
- Permissions apply to answers. A summary must not leak confidential source material.
- Deletion must remain possible. Preserve the path from source to knowledge entry, so that a deletion request reaches the entry.
- People decide standards. Frequency does not make a statement correct. Delegate approval to the AI because the expert owner has no time, and the collection holds what the model found plausible, until a customer notices.
A small first project
Choose ten approved conversations of one type. Extract no more than twenty recurring questions. Have an expert owner review the answers and publish them in a limited internal space first. Measure four things:
- Which questions are found?
- Where is the evidence insufficient?
- Which answers are corrected?
- Which entries are actually reused?
Then decide whether you need more sources, retrieval or a graph. The system grows from demonstrated demand instead of a tool demo. You do not have to choose the ten conversations alone: which conversation type brings the fastest return is regularly on the table in the calls of my community.
Where Gustav fits in the cycle
Gustav is my AI employee for the path from an approved conversation to a verifiable register, which means stages 3 and 6 and the role of the knowledge steward. Whisper runs locally. After the pseudonymisation pass, Gustav distils questions, objections, answers and source locations in batches of ten to twenty calls, with a sample review after every batch. He keeps the ledger that always shows which call belongs to which person, so a deletion request can be carried out. From this come sales objection cards, FAQs and the internal knowledge base, among other things. The expert owner remains the person who approves or rejects a statement. His most important rule is not a technical one: he distils what was said and invents no advice on top of it.
Frequently asked questions
What is AI knowledge management?
A cycle in which AI takes over collecting, clustering and comparing, and people decide what is valid. The six stages are source, rights, patterns, approval, return and review. AI makes the third and the sixth step faster, because those are the monotonous ones. Approval stays with an expert owner who has a name, and every answer carries a source and a validity date.
How do I capture expert knowledge that only exists in people's heads?
In conversation, not in a form. Someone who has been able to do something for years can rarely write it down, but can explain it. A handover interview in which a new colleague asks and the expert answers, recorded with her consent, delivers questions and answers in the language they will later be needed in. That becomes a register, and the entries where the expert says “I do that by feel” are the most important ones.
Which AI tools does knowledge management in a company need?
At first none that you buy. The first version is a register in a table or a note system, plus local transcription and a language model to pull out the questions. An AI knowledge base comes when colleagues are supposed to search for themselves, a knowledge graph when dependencies between entries start to matter.
What is the difference between knowledge management and a knowledge base?
Knowledge management is the process; the knowledge base is one of its tools. The process defines which sources count, who approves, when reviews happen and where the knowledge shows up in daily work. The knowledge base is the place where approved entries live and are found. A knowledge base without the process decays silently; the process without a knowledge base works as long as the team is small.
How often does knowledge have to be reviewed?
In my business quarterly, plus after every change to prices, offer or audience. After each conversation only collect, do not cluster. The quarterly run clusters, counts, rereads answers and closes open entries. That is a working rhythm from my own business, not a measured limit; for topics that change quickly it may be too coarse.
How to carry on
Answer one question: whose knowledge would the business miss if that person were out for three months? That is your first source. Arrange a one-hour handover interview with her, obtain consent, let a colleague ask the questions. From that one conversation, pull out the questions and their answers, and the entries where she says “by feel” are your starting point for the second interview.
The register is kept by Gustav in my business, and his package is in my community, with the pseudonymisation pass and consent templates. Who approves and on which rhythm is yours to define for your business, and you are not alone in that: in the calls we work through roles and review rhythms together, and a question in a post usually finds someone who has already organised the same handover. Everything about it is under Community.
