Building maintainable knowledge systems from messy business material

Files, email, decks, sheets, and meetings become a governed knowledge layer — with named owners, review dates, and provenance — so answers trace to sources instead of folklore.

8/29/2026

Every organisation has the same problem in different clothes: the truth exists, but it is scattered across files nobody can find, email nobody re-reads, decks with stale numbers, and meeting recordings nobody replays. AI makes retrieval cheap. It does not make the knowledge governed — that part is still your job.

Three stages: messy input from files, email, decks, sheets and meetings; extraction and structure with provenance; a governed layer with named owners and review dates
The pipeline is boring on purpose. That is why it holds up. Evidence date: 29 August 2026. Download the editable Excalidraw source.

The stack

Knowledge system stack: source corpus unchanged, extraction pipeline with provenance per claim, governed layer with owners and canonical versions, and retrieval for agents and people
Review dates are enforced. Stale entries are surfaced, not silently trusted. Evidence date: 29 August 2026. Download the editable Excalidraw source.

Source corpus stays raw. Files are never edited in place. The corpus is the evidence; the governed layer is the interpretation.

Extraction attaches provenance. Every extracted claim keeps its source: which file, which email, which meeting, which date. A claim without provenance is a rumour with formatting. (The delivery-side definition of provenance is owned by Proof, not prompts; this article owns the store.)

The governed layer is canonical. One version per fact, a named owner, a review date.

The non-obvious governance rule is the one that separates a knowledge layer from a search index: when two sources disagree, the layer records both and flags the conflict instead of silently picking one. This is the whole argument against naive retrieval — a system that auto-picks a winner launders the disagreement into a confident answer. Flagging the conflict keeps the human decision in the loop where it belongs.

The concrete instantiation on this rig is a docs tree with per-topic files, named owners, and enforced review dates, run weekly: the structure is boring, and that is the point.

Retrieval asks the layer. Agents and people query the governed layer, not the folder. This is the difference between "the AI said" and "the layer says, and here is the source".

Material to decision, in order

Five steps: material arrives, extracted, governed with canonical version, retrieved by agents, decision traceable to source
This sequence runs weekly on this rig. Evidence date: 29 August 2026. Download the editable Excalidraw source.

Folder versus governed layer

Before: folder of files with no owner per claim, no review date, contradictions invisible. After: governed layer with owners, review dates, surfaced contradictions, traceable answers
The governed layer catches what the folder silently hides. Evidence date: 29 August 2026. Download the editable Excalidraw source.

What AI changes and what it does not

AI makes extraction and retrieval nearly free. It does not decide who owns a fact, when it should be reviewed, or which of two contradictory sources wins. Those are governance decisions, and organisations that skip them get a fast way to be confidently wrong.

Where to go next