The relevant trust question is not whether AI appears in the product. It is who can request work, who must approve it, which permissions apply at execution and what evidence remains.
Content reviewed
Human control is a workflow, not a disclaimer
A proposed action has an owner and state. Approval does not bypass current authorization, external agent access is constrained by capabilities and consequential changes pass through the platform audit path.
Trust topic
Follow the decision from proposal to evidence
These controls describe the implemented orchestration path. They do not claim that generated content is always correct or that every product action is AI-assisted.
01
How does a proposed action remain owned by a person?
Agent actions are represented as proposals with lifecycle state and user-facing review controls. Consequential work waits for an explicit person-led decision before execution.
What is implemented
The action data service persists proposal state; the agent workspace presents review and approval interactions to the operator.
Boundary
A human approval gate reduces silent execution risk; it does not guarantee the person will make a correct decision. Organizations still need appropriate reviewers and procedures.
02
Does approval override the user's current permissions?
Execution authorization is evaluated against current user and organization context. A previously created proposal is not a standing grant to bypass later permission changes.
What is implemented
The authorization service resolves execution rights, and request authentication supplies current identity and effective permissions.
Boundary
This control depends on the action using the central authorization path. New tools require the same integration and review before being treated as covered.
03
How is external agent access constrained?
External agent connections enter through the platform's MCP route and operate against registered capabilities rather than receiving unrestricted application or database access.
What is implemented
The MCP route validates connection and tool context; the capability schema records the available action contract and its governance metadata.
Boundary
Capability registration is not permission by itself. Effective authorization and the action's own validation still determine whether a request may proceed.
04
What evidence remains after a consequential action?
Business audit events pass through a central recording service and can be exposed in the organization audit feed with actor, action, target and timing context.
What is implemented
The audit data path records events and the agency audit-feed module provides the organization-facing read surface.
Boundary
Audit usefulness depends on every consequential feature emitting the right event. Coverage must be reviewed as new modules and actions are introduced.
Question to ask
Questions for a consequential-action review
Use an action your organization cares about and inspect proposal ownership, approval, permission changes and the resulting history.
01
Where is the irreversible or consequential decision?
Identify the exact moment that changes a record, sends something, spends money or affects a client. Confirm who approves it.
02
Whose permissions are checked at execution?
Test a proposal before and after changing the approver's permission. Confirm that ownership and current authorization remain distinct.
03
What can an administrator review afterward?
Inspect proposal state, actor identity, approval, execution result and the business audit event for the scenario.
Related control paths
Continue with the security overview, responsible-assistance explanation and the knowledge-and-control capability ledger.
