Evaluating company brain products

Buying (or building) a company brain is easy to get wrong if you score demos on chat quality alone. Fluent answers are cheap. Governed memory is the product. Use this checklist to compare vendors and internal builds without fabricated ROI theater.

1. Ownership and provenance

Ask for a concrete demo:

  • Assign an owner to a record; show it in the UI and API
  • Show source links and change history for a claim
  • Show who last verified a fact and when

If the system only stores anonymous chunks with embedding timestamps, you are buying search, not a brain.

2. Freshness and drift

Require:

  • Review / expiry signals on high-stakes records
  • Conflict handling when two sources disagree
  • Metrics or views for stale content touching agent traffic

“We reindex nightly” is infrastructure. Drift control is product.

3. Access control for people and agents

Probe:

  • Can agent identities have narrower scopes than the humans who invoke them?
  • Can you deny write/promote while allowing read?
  • Do permissions map to your IdP groups and data classes?
  • Are memory accesses logged for audit?

See Agent memory access control for the deeper pattern.

4. Agent interfaces

A company brain should expose more than a web search box:

  • APIs or tool/MCP interfaces with structured read/write
  • Citations suitable for agent traces
  • Clear failure modes when policy blocks retrieval

If “integration” means “paste into ChatGPT,” the brain is not in the loop.

5. Write and promotion paths

Look for:

  • Human review queues for promoting tribal knowledge
  • Separation between draft memory and canonical memory
  • Guards against agents silently polluting the corpus

6. Portability and lock-in

Prefer:

  • Export of records in open formats
  • Git-friendly or otherwise versionable knowledge where it fits (Git-owned company knowledge)
  • Clear data residency and deletion story

Escape hatches matter more than launch-day UI polish.

7. Fit to systems of record

The brain should explain and connect, not pretend to replace CRM, code, or billing. Ask how live system data is referenced vs duplicated, and what happens when the CRM changes.

Scoring without vanity metrics

Weight categories by your risk profile. A regulated support agent needs permissions and audit more than a brainstorming wiki. An eng-heavy org may weight git ownership higher. Record scores in a one-pager your team can revisit after a pilot.

Where Mensara fits

Mensara is a product in this category focused on governed company memory for AI-native teams. Evaluate it with the same checklist as anyone else: owners, freshness, policy, agent interfaces, and export — not slogan proximity.

Pilot design

A useful pilot is narrow:

  1. Pick 20–50 high-traffic questions or playbooks
  2. Assign owners; load governed records
  3. Route one agent path through the brain with logging
  4. Measure contradiction rate, stale hits, and time-to-correct — not “user delight” alone

Continue with What is a company brain? and Governed memory vs RAG if your team still conflates retrieval with governance.