Knowledge Management: Workflows
How knowledge is captured, made authoritative, retrieved and retired, and why this function is load-bearing in an AI-operated company.
Markdown. No sign-up, no email.
In a conventional company this function is administrative. In one where agents read the documents to decide how to act, it is load-bearing, and a stale page stops being an inconvenience and becomes an instruction.
What arrives#
| From | What | Becomes |
|---|---|---|
| Every function | Decisions and their reasoning | Decision records, within 24 hours |
| Engineering | Architecture decisions, and rejected options | The record that stops re-argument |
| Security and DevOps | Postmortems | Lessons, and detection improvements |
| Research | Findings, especially negative ones | The record that stops paying twice |
| Sales | Proposals, and what won | Reusable material |
| PMO | Project history and client context | Departure-proof knowledge |
| People leaving | What only they know | Captured in the first week of notice, not the last |
What leaves#
| To | What |
|---|---|
| Anyone asking | An answer with citations, or an honest "not here" |
| Every agent | The context bundle it reads, reviewed quarterly |
| New joiners | A path from nothing to useful |
| Marketing | Material worth publishing, once cleared |
| Leadership | The knowledge risk register: what exists in one head only |
The cycle#
The third lane is the one nobody builds, and it is the difference between a knowledge base and an archive.
Handoff contracts#
With every function. Decisions arrive within 24 hours with their reasoning. The agent drafts; the decision owner confirms. Nobody is asked to write a document from scratch, because that is the requirement that quietly ends every knowledge programme.
With Engineering. Architecture decisions arrive with rejected options. Without them, the same debate recurs every eighteen months among people who cannot tell whether the current design was chosen or defaulted into.
With HR. Departure capture starts in the first week of a notice period.
With AI Engineering and Product. What each agent reads is agreed here and reviewed quarterly. It is the highest-consequence text in the company and the easiest to let accumulate.
Cadence#
| Continuous | Capture |
| Monthly | Staleness, duplicate authorities |
| Quarterly | Retirement pass, context review, knowledge risk register |
| On notice | Departure capture, first week |
The failure this design is built against#
A wiki with four thousand pages, of which perhaps eight hundred are still true, and nobody able to say which. People stop using it, ask a colleague instead, and the knowledge that was supposed to be institutional is once again personal.
In an AI-operated company that failure escalates: the agents cannot ask a colleague. They read what is there, believe it, and produce confident, consistent, incorrect work at volume.
Retirement is the whole answer, and it is why the retirement rate is a headline metric rather than housekeeping.