For years, knowledge management meant a wiki that slowly went out of date. Agents change the stakes. An assistant that can read your documents and answer questions will repeat what it finds, including what is stale, contradictory or wrong. At the same time, agents make good knowledge far more valuable, because every correct page now helps every person and every agent that asks. The organizations that treat knowledge as a product with owners will get more from AI than those that treat it as an archive.
The three kinds of knowledge
- Explicit knowledge is already written: policies, manuals, specifications, tickets. It is the easiest to connect to an agent and the most likely to be duplicated and outdated.
- Tacit knowledge lives in people's heads: how things really get done, why a decision was made, what to watch out for. It is the most valuable and the least captured.
- Contextual knowledge is the situation around a question: the customer, the system, the history, the constraints. Agents need context to give an answer that fits.
Step one: capture what people know
The richest knowledge appears in retrospectives, incident reviews, onboarding conversations, cohort sessions and the questions colleagues ask each other. Build small, regular capture habits into these events. A short template works: the situation, what was tried, what happened, what we would do next time and when this applies. Agents can help by drafting the note from the notes or recording of a session, but a person should confirm it before it is published.
Step two: structure it for retrieval
- Write for a stranger. Spell out terms and abbreviations. A reader, human or machine, will not share your hallway context.
- One idea per item. Short, focused entries retrieve better than long mixed pages.
- Add metadata. Title, owner, date, audience, the system or process it applies to and a review date.
- Link related items so that an answer can point to deeper material.
- Mark the source of truth. When two documents disagree, one should be clearly authoritative and the other retired.
Step three: assign owners and review dates
Every item that agents rely on should have a named owner and a date by which it will be reviewed. When the date passes, the item is flagged, and an agent can surface a short list of overdue items to each owner every month. This is the single most effective defense against stale answers. See the controls described in our AI and data governance guide.
Step four: ground the agents
- Limit the sources. Let an agent answer from specific, approved collections rather than everything.
- Show the source. Answers should cite the item they used, so people can check and so errors can be traced to a page.
- Allow “I don't know”. An agent that admits a gap is more useful than one that guesses. Gaps are a to-do list for capture.
- Respect access. An agent should not reveal content to someone who could not open it themselves.
- Collect feedback. A thumbs-down with a reason is a signal to fix an item.
A feedback loop that keeps knowledge alive
Treat questions people ask agents as data about what the organization needs to know. Review them monthly. Questions that get good answers show what is working. Questions that get poor answers or none show where to capture next. Share the best new items in your newsletter or team channel so people see that contributing matters. That is the capture-and-apply half of the Seek, Share, Capture, Apply loop.
Metrics worth watching
- How many questions are answered with a cited source and how many end in “not found”.
- How many items are past their review date.
- How long a new colleague takes to find the answer to a common question.
- How many new lessons were contributed this month, and by how many different people.
We offer no benchmark for these. Track your own trend over a few months and look for movement in the right direction.
Where to start
Choose one team and one body of knowledge, such as the onboarding guide or the support runbook. Clean it, assign owners, add review dates, connect it to an assistant and invite ten colleagues to use it for a month. Then decide, based on what they asked, what to capture next. Our book Context-Driven goes deeper on building shared context, and our AI learning agents page shows how agents use your organization's own knowledge. The learning organization guide covers the habits behind it.