Where human review belongs
Human review is not a disclaimer at the bottom of a screen. It is a design decision about which actions a system may take alone, which it must propose, and what the reviewer needs in order to decide quickly.
Give the reviewer a decision
A review queue works when the reviewer is asked one clear question, shown the evidence behind the answer, and given a way to correct it that is faster than doing the work again. If reviewing costs more than redoing, the queue gets ignored.
What belongs in front of a reviewer
- The proposed action, stated plainly, with a default the reviewer can accept.
- The sources behind it, linked, so a claim can be checked in seconds.
- The rule or precedent applied, so a disagreement can be traced to a cause.
- An edit path, so correcting the output is one action rather than a rewrite.
- A record of what was approved, edited, or rejected, and by whom.
Where review is not optional
Anything that reaches a customer under the company's name.
- Anything that changes a commitment, a price, or a legal position.
- Decisions about people: hiring, termination, eligibility, or access.
- Any action that is expensive or slow to reverse once taken.
“Every correction is a test case. A review queue that captures why the reviewer disagreed is worth more than one that only records that they did.”
Conclusion
Autonomy should be earned in public. Start with review on everything, watch where reviewers accept without changes, and move those cases to automatic. The queue then holds the work that genuinely needs judgment.
