Human-in-the-Loop Approvals

Approval should answer: who/what is asking, what will happen, why it is risky, and what I am authorizing.

Priority: P0
Phase: MVP
Primary owner: Workflow / Security
Status: Blueprint

Objective

Human approval is the bridge between autonomous speed and accountable high-impact actions. The workflow should provide the approver enough context to make a decision quickly without forcing them to open several systems.

What to build

Approval triggers by action type, amount, data class, risk score, destination or policy.
Approver routing by manager, application owner, security team, finance role or custom group.
One-click approve/reject with reason, expiry and optional parameter modification.
Escalation, timeout and fallback behavior for unattended approvals.
Approvals through web, email, Teams/Slack connectors where permitted.

Implementation decisions

Display agent identity, principal, requested action, target, parameters, policy reason, evidence and expected consequence.
Use signed single-use approval tokens and prevent replay.
Support four-eyes approvals for critical actions.
Make approval SLA part of policy to avoid indefinite queued actions.

Definition of done

A high-risk action cannot execute before required approval is cryptographically bound to the request.
Audit record includes who approved, what exact request and which policy required it.
Expired/revoked approvals cannot be reused.

Success metrics

Approval response time
Approval/rejection rate
Expired requests
Critical actions with dual approval

UX RULE

Approval should answer: who/what is asking, what will happen, why it is risky, and what I am authorizing.

This control is specified on page 26 of the blueprint.