Knowledge Management: Charter
The function that makes the company's experience reusable, owning one authority per topic, the retirement of stale documents, and the context every agent is given to work from.
Knowledge management is usually where good intentions go to be archived. A wiki is created, everyone contributes for a month, and two years later nobody trusts a page because nobody knows which of the four versions is current.
This function exists to prevent that, and in an AI-operated company it stops being administrative and becomes load-bearing: the documents are what the agents read. A stale page is no longer just misleading to a colleague: it is an instruction, followed at scale, without hesitation.
What this function owns#
One authority per topic. Exactly one current document per subject, and everything else marked superseded with a pointer. Duplication is the failure mode, not scarcity.
Retirement. Actively removing what is no longer true. This is harder than writing and matters more.
The agent context layer. Which documents are loaded into which agent's working context. No text in the company changes more outputs per word, and it must not accumulate by accident.
Findability. Search, structure and naming: measured by whether people find the answer, not by whether it exists.
Capture at the point of work. Decisions and lessons recorded when they happen, because a retrospective written a month later records the story rather than the reasoning.
What is NOT delegated to an agent#
- Declaring which document is the authority when two conflict.
- Retiring a document. Deletion is irreversible in effect even when it is reversible in a version history.
- Deciding what goes into an agent's context. That shapes every output that agent produces.
- Recording a decision's reasoning. An agent can draft the summary; the person who made the call confirms why.
KPIs#
| Measure | Why this one |
|---|---|
| Median document age at last review | Rising age is the leading indicator of a rotting knowledge base |
| Duplicate-authority incidents | Two current documents on one topic is a defect, counted as one |
| Time to find an answer | Sampled by asking people, not inferred from search logs |
| Share of decisions with a recorded rationale | The reasoning is what stops the same debate recurring |
| Superseded documents still reachable without a pointer | Each one is a trap |
| Context-layer size per agent | Grows silently, costs money, and dilutes what matters |
AI agents in this function#
Staleness agent. Flags documents whose review date has passed or whose referenced systems no longer exist. Read-only; it proposes, a person retires.
Duplication agent. Detects two documents claiming authority on one topic and raises a conflict rather than merging them silently.
Capture agent. Drafts a decision record from a meeting or a thread, which the decision owner then corrects and confirms. The draft removes the excuse; the confirmation keeps it honest.
Context-audit agent. Reports what each agent is actually being given to read, and what in that bundle has never influenced an output. Unused context is cost and noise.
Consistent with the rest of the framework: agents surface, humans decide. Nothing is retired, merged or promoted to authority by an agent.
SOPs#
- One-authority rule. A new document on an existing topic either replaces the current one or does not ship.
- Review dates on everything. A document without one is treated as expired.
- Supersede, do not delete. The old version stays reachable with a pointer forward, so an old link teaches rather than misleads.
- Decision records captured within 24 hours, while the reasoning is still recoverable.
- Quarterly context review. What each agent reads, and why.
Templates#
SOP, Business Requirements, decision record, lessons-learned record, document review log.
Workflows#
In: decisions from every function, incident reviews from Operations and Security, negative results from Research & Innovation, process changes from AI Strategy & Transformation.
Out: the current authority set, agent context bundles, onboarding material, the searchable record of why things are the way they are.
Handoffs: HR for onboarding paths, AI Engineering & Product for anything loaded at runtime, Compliance for documents with a retention obligation.
The loop: capture → publish as the single authority → review on a date → supersede. Break the review step and the base decays invisibly, which is the one failure nobody notices until an agent acts on it.
FAQ#
Why does this matter more in an AI-operated company?#
Because documents stop being reference material and become instructions. A human reading an out-of-date page often senses something is wrong; an agent reads it, believes it, and produces consistent, confident, incorrect work at volume.
Is a wiki enough?#
Only with retirement. A wiki without an owner and review dates becomes an archive of things that were once true, and its search results actively mislead because everything looks equally current.
What should be captured?#
Decisions and their reasoning, incidents and their causes, and negative results. Not status updates, and not anything a system already records, duplicating those creates a second version to disagree with the first.
Who owns a document?#
A named person, not a team. Team ownership means review dates pass without anyone feeling responsible, which is how the base rots while appearing maintained.
What else is coming for Knowledge Management
Charter Ready
What this department owns and is accountable for.
KPIs Not yet
The numbers it is judged on.
AI Agents Not yet
What is automated, and what stays human.
SOPs Not yet
How the recurring work is done.
Templates Not yet
The documents it produces.
Workflows Not yet
How work enters, moves and leaves.