Human-In-The-Loop Approval Matrix
Use this when deciding what humans approve, sample, audit, or own in an AI workflow.
Workflow Facts
- Workflow name:
- Owner:
- Users affected:
- Customer-visible? Yes / No
- External action? Yes / No
- Sensitive data involved:
- Reversibility:
- Expected volume:
- Exception owner:
Consequence Matrix
| Question | Yes / No | Notes |
|---|---|---|
| Can a wrong output affect a customer, employee, supplier, or regulator? | ||
| Can the workflow send, publish, delete, charge, refund, or change records? | ||
| Can it expose personal, confidential, financial, legal, HR, or security data? | ||
| Would the mistake be hard to detect later? | ||
| Would the mistake harm trust even if reversible? |
Review Pattern
| Consequence | Default pattern | Selected? |
|---|---|---|
| Internal, reversible, low visibility | Audit after the fact | |
| Repeated, quality-sensitive work | Sample review | |
| Routine flow with ambiguous cases | Exception review | |
| Customer-visible or external action | Human approval before action | |
| Destructive, financial, legal, HR, or regulated decision | Human owns final decision |
Review Screen Requirements
- Proposed output or action is visible.
- Source evidence is visible.
- Risk flags are visible.
- Reason for review is visible.
- Reviewer can approve, edit, reject, or escalate.
- Final action is logged with reviewer and timestamp.
Stop Rules
- Error-rate threshold:
- Queue-age threshold:
- High-risk incident threshold:
- Owner who can pause workflow:
- Manual fallback path: