Markdown template
Company AI context readiness checklist
A practical checklist for testing shared company knowledge across people and AI agents without weakening identity, access, review, or tenant boundaries.
A company AI context pilot is ready to expand only when approved sources and changing continuation records are separated, every person and agent has a distinct identity, search and writes enforce the same access rules, review links to immutable revisions, revocation is tested, and a fresh actor can continue without the original chat.
Best for: Company administrators, platform teams, and multi-agent pilot owners
Updated August 4, 2026
Download MarkdownVerification workflow
Company AI context readiness workflow
Complete these checks and preserve the evidence before expanding shared context across the company.
- 01
Inventory approved sources and canonical owners
Record which documents, datasets, and systems may ground agent work and who can approve changes.
- 02
Separate source knowledge from continuation records
Keep stable reference material distinct from changing objectives, decisions, evidence, review, and next actions.
- 03
Give every person and agent a distinct identity
Use SSO for people and single-purpose service identities for unattended agents; never accept authorship from prompt text.
- 04
Declare company, workspace, and record boundaries
Define tenant isolation, memberships, roles, scopes, visibility, and the limited organizational role of folders.
- 05
Make writes attributable and revisioned
Preserve immutable evidence and create a new revision or collaboration event for every material change.
- 06
Apply authorization before search and retrieval
Filter every index and direct read by the active principal so metadata cannot reveal denied records.
- 07
Test review, revocation, retention, and recovery
Prove corrections link to revisions, disabled identities lose access, retained history survives, and recovery is documented.
- 08
Run a two-person, two-agent continuation pilot
Require correct visibility and authorship plus one successful fresh-session continuation before expanding company-wide.
The company context layer is not ready when...
- One global agent token or copied human token represents several actors.
- The search index reveals titles, snippets, or counts from unauthorized workspaces.
- Folders are treated as security boundaries without explicit authorization rules.
- Generated summaries replace canonical source files or immutable revisions.
- A mention, copied link, or matching email domain silently widens access.
- The pilot counts imported items but does not test review, revocation, and continuation.
Included fields
What this template preserves
- Sources
- Identity
- Boundaries
- Write model
- Review
- Search
- Lifecycle
- Pilot result
# Company AI context readiness checklist
Use this record before expanding a shared knowledge base across people and AI
agents. Do not include credentials, tokens, secrets, or personal data that
belongs in a protected system.
## Readiness checklist
- [ ] Approved sources and their canonical owners are inventoried.
- [ ] Source knowledge is separate from changing continuation records.
- [ ] Every person and unattended agent has a distinct identity.
- [ ] Company, workspace, and record-level access rules are explicit.
- [ ] Every material write creates attributable, revisioned evidence.
- [ ] Search and direct reads enforce the same authorization checks.
- [ ] Review, revocation, retention, and recovery have been tested.
- [ ] A two-person, two-agent pilot passed the continuation test.
## Approved sources
| Source | Canonical owner | Classification | Allowed workspaces | Refresh method |
| --- | --- | --- | --- | --- |
| | | | | |
## Context layers
### Source knowledge
<!-- List stable policies, specifications, datasets, and reference material. -->
### Continuation records
<!-- List changing objectives, decisions, evidence, review, and next actions. -->
### Derived indexes
<!-- State how search points back to canonical, permitted records. -->
### Private session memory
<!-- State what must never be treated as company evidence automatically. -->
## Identity inventory
| Principal | Type | Human owner | Purpose | Workspaces | Allowed actions |
| --- | --- | --- | --- | --- | --- |
| | Person / service agent | | | | |
## Boundary map
- Company:
- Workspaces:
- Folder purpose:
- Record visibility options:
- Public-sharing owner:
- Explicitly denied paths:
## Write and review model
- Canonical record:
- Revision rule:
- Artifact integrity rule:
- Annotation and assignment rule:
- Resolution evidence:
- Conflict handling:
## Search and retrieval policy
- Indexed fields:
- Authorization filter:
- Direct-read parity check:
- Metadata leakage test:
- Canonical result link:
## Lifecycle and recovery
- Access review cadence:
- Disable and revocation procedure:
- Retention policy:
- Export or backup procedure:
- Recovery test:
- Offboarding owner:
## Pilot verification
- [ ] Two separately authenticated people participated.
- [ ] Two separately authenticated agents participated.
- [ ] Each actor saw only permitted sources and records.
- [ ] Every revision showed the authenticated author.
- [ ] A human finding linked to a correcting revision.
- [ ] A denied company or workspace search returned no metadata.
- [ ] A disabled agent immediately lost read and write access.
- [ ] A fresh actor completed one bounded next action without the original chat.
## Pilot result
**Status:** Pass / Blocked
**Pilot owner:**
**Evidence links:**
**Blockers:**
**Next review date:**
Questions
Company AI context FAQ
What is a shared knowledge base for AI agents?
It is a company-scoped system where authorized people and agent identities can retrieve approved sources and contribute durable, attributable work under the same access rules.
Is a vector database enough for company AI context?
A vector index can improve retrieval, but it does not by itself preserve canonical files, identity, revisions, review, visibility, retention, or continuation ownership.
Should all company agents share one memory?
No. Agents may use the same governed knowledge layer, but each identity should see only its permitted workspaces and records. Private session memory should not become company evidence automatically.
How do you test tenant isolation for AI agents?
Use separately authenticated identities in different companies and workspaces, then verify that search, direct reads, artifacts, annotations, and writes expose only explicitly permitted records.