Knowledge Management: KPIs
Retrieval success, document currency and the retirement rate, measured because a knowledge base grows by default and only shrinks on purpose.
Markdown. No sign-up, no email.
Every measure here is chosen to resist the same failure: a knowledge base that grows continuously, is never pruned, and slowly becomes an archive of things that were once true.
The six that matter#
| Measure | Definition | Target | How it gets gamed |
|---|---|---|---|
| Retrieval success | Questions answered from the Brain with correct sources | Above 80% | Counting an answer as success because one was produced |
| Median document age at last review | Across everything current | Under 12 months | Re-dating without re-reading |
| Retirement rate | Documents superseded or removed per quarter | Above zero, always | Never removing anything, which is the default |
| Duplicate authorities | Topics with more than one current document | Zero | Not looking. Nobody trips over this until an incident |
| Decisions with recorded reasoning | Material decisions carrying a why | Above 90% | Recording the decision and omitting the rejected options |
| Time to find an answer | Sampled by asking people | Under 5 minutes | Inferring it from search logs, which measure searching rather than finding |
Retirement rate above zero is the health check for the whole function. A knowledge base that only grows is one where the share of true content falls every quarter. Nobody notices, because nothing visibly breaks until someone acts on a page from 2024.
Time to find an answer is sampled by asking people, not derived from analytics. Search logs tell you what was searched, not whether the person found what they needed or gave up and asked a colleague.
Two counter-metrics#
| Counter-metric | Catches |
|---|---|
| Questions answered by asking a person instead of the Brain | The honest measure of whether this works. If people still ask a colleague first, the Brain is not trusted |
| Context bundle size per agent | Grows silently, costs money, and dilutes what matters. Nobody prunes an agent's reading list unless it is measured |
Knowledge risk, reviewed quarterly#
| Question | Bad answer |
|---|---|
| Which knowledge exists only in one person's head? | Any critical system or client |
| Which clients have no written context? | Any. This is what makes a departure expensive |
| Which architecture decisions have no recorded rejected options? | Most, usually. That is where re-argument comes from |
| What would we lose if the longest-serving person left tomorrow? | If this cannot be answered, the answer is "a lot" |
What is deliberately not measured#
- Documents created. Growth is the failure mode, not the goal.
- Wiki page views. Reading is not finding.
- Contribution counts. Rewards volume over authority, and volume is what breaks this.
- Search volume. High search with low retrieval success means people are hunting, not finding.
Cadence#
| Continuous | Capture at the point of work |
| Monthly | Staleness flags, duplicate authorities |
| Quarterly | Retirement pass, context bundle review, knowledge risk register |
| Annually | Whether the structure still matches how people actually look for things |