CHECK 01 / ACCESSYOUR EXISTING CONTROLS
Is the agent allowed to act?
Access controls answer who can use which tool—and within what limits.
Identity verifiedTool access allowedAction scope in bounds
Access control.
Plus the check it misses.
Identity, tool access and action scope are table stakes for any agent platform. Company Brain adds the control access platforms cannot see: whether the knowledge behind an already-authorised action is still current and approved.
01
THE TWO CONTROLS
TWO SEPARATE JOBS, RUN TOGETHER
Who may act.
What they may act on.
Access control decides whether an agent is authorised. Knowledge control decides whether what it is about to say is still true. Company Brain runs the second job alongside your existing access-control platform—not instead of it.
ONE AGENT / ONE DECISION“Can I approve this refund?”REQUEST: $750
CHECK 02 / KNOWLEDGECOMPANY BRAIN
Is the knowledge safe to use?
Knowledge controls answer whether the guidance is current, owned and released for use.
Source v18.2 currentOwner approvedAnswer impact released
CHECK 03 / EXECUTIONCOMPANY BRAIN
Is the intelligence allowed to run this workload?
Execution controls answer whether the model and provider are approved for this workload, inside its cost and quality boundary.
Model approvedProvider in policyCost / quality in bounds
ONLY WHEN ALL ARE TRUEAccountable AI.
The agent has permission, current knowledge and an approved execution path.
✓Customer-controlled environment
One isolated production environment is deployed into infrastructure you provide and control. In the standard configuration, document content and inference stay inside that environment; a premium hosted model is optional and off unless you enable it.
Credentials encrypted, keys durable
Source credentials are encrypted before they reach the database, with cryptographic separation per source. Production start fails closed rather than use a key ring that could leave credentials unrecoverable.
Tamper-evident operating record
Governance-critical changes are written to an append-only, hash-chained record. The chain can be verified on demand and reports the sequence where a break is detected.
Human release decision, enforced in code
No knowledge reaches a published state without cited, hash-verified evidence and a named reviewer's decision. The deterministic grounding checks sit on one release path.
Responsibility lines, engineering evidence
and what is not yet claimed.
The deployment boundary, internal test evidence and an explicit list of what is not certified live on the full security and deployment page—including what Company Brain manages, what the customer still controls, and where independent certification is still planned rather than held.
See security & deployment →