Human review is a system property
A reviewer can appear in a workflow diagram and still have no meaningful control. If the person receives a polished recommendation without sources, time to challenge it, or the ability to prevent downstream action, the review is ceremonial.
The real test is whether the operating design makes disagreement possible and consequential.
The seven-control model
- Accountable owner — name the role responsible for the decision and review quality.
- Evidence packet — present sources, dates, jurisdiction, missing facts, and the draft together.
- Decision boundary — state what the system may prepare and what it may never approve or execute.
- Stop conditions — force refusal or escalation for sensitive topics, conflicts, and confidence gaps.
- Permission boundary — prevent the agent from bypassing the review through another tool or action path.
- Recourse path — give the affected person or downstream user a route to a qualified human.
- Decision record — capture the version, reviewer, changes, rationale, owner, and completion evidence.
Separate facts, interpretation, and action
A reviewable artifact should show source facts, scope, interpretation, missing context, proposed next step, and accountable owner as separate layers. That structure helps the reviewer confirm the evidence and see where qualified HR, legal, privacy, or security review is required.
A minimum acceptance test
- The system refuses a sensitive named-employee decision
- Every material claim links to a source
- Missing facts appear before any proposed action
- The reviewer can edit, reject, or reroute
- No consequential action occurs before approval
- The record shows what changed between draft and final
If the human cannot see why the system produced the output, cannot stop it, or cannot leave a durable correction, the human is supervising in name only.