CORE GUIDE
Human review boundaries
Human review is useful when it owns a concrete decision at a consequential boundary, not when a person merely glances at AI output.
Mental model
A human review boundary is a policy-controlled handoff: the system defines which actions require review, what evidence the reviewer receives, what decision they must make, and what happens after approve, reject or request-changes.
Why it matters
Adding a generic human-in-the-loop step can create the appearance of safety without meaningful control. Reviewers become rubber stamps when they see too much low-value context, cannot inspect the relevant evidence, or are not empowered to block an action. Strong review boundaries are reserved for decisions where impact, irreversibility, policy ambiguity or accountability justify human judgment.
01
Design the handoff as a decision contract
Classify actions by impact and reversibility, then define escalation triggers such as policy conflict, missing evidence, high monetary value or destructive side effects. Package the smallest evidence set needed for the reviewer to decide. The runtime must pause the consequential action until an explicit review outcome is recorded, and that outcome should become durable audit evidence.
02
Example: support agent refund escalation
A support agent can search policy and draft replies automatically. Small credits within a predefined rule are allowed, but a large refund or conflicting policy evidence creates a review task containing the customer state, controlling policy, proposed action and reason for escalation. The reviewer approves, edits or rejects the action; the model cannot bypass the decision by rephrasing the request.
Common failure modes
- Putting human review after the irreversible action has already occurred.
- Showing reviewers a full transcript instead of the evidence needed for the decision.
- Using vague escalation language without machine-enforced triggers.
Engineering heuristics
- Define the exact decision the human owns and the actions that must wait.
- Escalate based on impact, ambiguity and reversibility rather than model confidence alone.
- Record review outcomes as durable runtime state and audit evidence.
Takeaways
- 01Human-in-the-loop is a boundary design problem, not a checkbox.
- 02Review is strongest before consequential side effects.
- 03A reviewer needs decision-relevant evidence and real authority to block.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.