Llm wiki company flow audit

LLM-Wiki Company Flow Audit

Goal

Help users plan company/team-level LLM-Wiki ingestion and operations from real information flows, without pretending that all knowledge can be safely or usefully automated.

When to use

Use when the user asks:

• where team/company information currently lives;
• how to collect from Confluence, Jira, GitHub, Slack, Teams, Drive, meetings, support, analytics or email;
• what can be fully automated and what cannot;
• how to keep external docs updated when they change;
• how fragile vault-review/lint systems are;
• how much time link/contradiction repair takes;
• whether there is a good UI for review queues;
• whether confidential information and access control are supported;
• what pitfalls to avoid;
• whether the maintenance cost is worth the value.

Inputs

• The list of systems currently holding team/company knowledge (e.g. Confluence, Jira, GitHub, Slack, Teams, Drive, PagerDuty).
• Team/org scale (personal pilot, one team, cross-functional, or company-level) so the maintenance-planning range fits.
• Whether a source registry or ingestion tooling already exists, and its refresh policy (webhook, poll, manual, snapshot).
• Any confidentiality/access constraints: data classification tiers, who owns which sources, existing permission boundaries.
• Prior review-queue or UI setup (PR-based review, Obsidian vault, dashboard) if one is already in place.

Required references

Read these when available:

• references/docs/company-information-flows.md
• references/docs/adoption-objections.md
• references/docs/adoption-q-and-a.md
• references/docs/19-security-threat-model.md

For vendor claims about Confluence, Jira, Slack, Teams, Google Drive, SharePoint or GitHub, browse and cite current official docs.

Core stance

Use this answer:

Start from the information flows that already influence decisions. Automate capture, metadata, version/hash tracking, stale marking and review reports. Do not fully automate truth promotion, contradiction resolution, sensitive summaries or official documentation updates.

Procedure

1. Inventory source flows

Map sources into:

Flow	Examples
Docs	Confluence, Notion, Google Docs, SharePoint, PDFs.
Work tracking	Jira, Linear, GitHub Issues, YouTrack.
Code collaboration	PRs, reviews, commits, releases, ADRs.
Chat	Slack, Teams, Telegram, Discord.
Meetings	Calendar, transcripts, recordings, notes.
Customer/product evidence	CRM, support tickets, sales calls, research repos.
Analytics	BI, warehouse queries, product analytics, experiments.
Ops/incidents	PagerDuty, postmortems, on-call notes.

Recommend starting with 3-5 flows that affect real decisions, not every system.

2. Classify automation level

Use three levels:

Level	Meaning	Examples
Automatic capture	Safe append-only or metadata sync.	Webhooks, event logs, source hashes, stale flags.
Assisted transformation	Agent drafts structure but review is needed.	Summaries, entity pages, decision pages, dedup candidates.
Human-gated truth	Human owner must approve.	Verified claims, contradictions, sensitive summaries, official docs.

3. Design external update handling

Require a source registry:

source_id: confluence:123456
system: confluence|jira|github|slack|teams|drive|manual
external_id: "123456"
url: "https://..."
version: "42"
updated_at: "2026-07-07T10:00:00Z"
source_hash: "sha256:..."
last_seen_at: "2026-07-07T10:05:00Z"
last_ingested_at: "2026-07-07T10:10:00Z"
derived_pages: []
refresh_policy: webhook|poll|manual|snapshot

Update flow:

webhook/delta/poll -> raw event -> fetch current source -> compare version/hash -> mark derived pages stale -> create diff/refresh report -> owner review

Do not silently overwrite reviewed human synthesis.

4. Explain fragility and repair effort

Be honest:

• broken links and orphans are usually cheap;
• schema drift is manageable with validators and dry-run migrations;
• duplicates and contradictions require owner judgment;
• stale external sources require a source registry and refresh loop;
• permission leaks and generated slop are high-risk;
• ignored review queues are a system failure, not a UI bug.

Use planning ranges, not guarantees:

Scale	Maintenance planning range
Personal/tiny pilot	30-60 minutes weekly.
One team/domain	1-2 hours weekly plus domain-owner review.
Cross-functional pilot	Product Ops/curator rotation and weekly review meeting.
Company-level Nexus	Dedicated ownership, access policy and dashboards.

5. Discuss UI

Default UI stack:

• Markdown reports for portability;
• GitHub/GitLab PR or issue queue for review;
• Obsidian graph/backlinks/Dataview for local exploration;
• Jira/Linear board for team triage;
• dashboard only after queues are too hard to operate manually.

Do not imply a polished universal UI exists unless the repository/tooling actually has one.

6. Handle confidentiality

Say:

Yes, but only with data classification, split access domains, least-privilege agents, model boundaries, redaction and audit logs. If permissions cannot be modeled, keep metadata/links only and fetch with user-scoped credentials on demand.

Never recommend ingesting secrets.

7. Evaluate cost versus benefit

Do not use page count. Use:

• time-to-answer recurring questions;
• answer reuse rate;
• decision citation rate;
• stale/unsupported claims found before decisions;
• review backlog;
• output beyond wiki: PRDs, strategy notes, experiments, ADRs, reports.

Continue only if the wiki changes decisions, speeds context recovery or prevents mistakes.

Output

Use this structure:

## Information-flow map

## Automation plan

## External update / versioning plan

## Fragility and review effort

## UI and operating model

## Confidentiality and permissions

## Pitfalls

## Cost-benefit check

Safety gates

• Do not promise full automation of truth.
• Do not claim webhook delivery is sufficient without reconciliation.
• Do not recommend all-chat or all-email ingestion by default.
• Do not centralize confidential sources before modeling permissions.
• Do not imply a polished UI exists unless it does.
• Do not use page count as a success metric.
• Do not ingest secrets.