Human-in-the-loop gates for AI agents: when to stop and ask

<p>Human review keeps outcomes safe without turning automation back into manual work. Place a gate only where a mistake is expensive.</p><h2>Worked example</h2><p>An agent drafts outbound messages and updates the CRM.</p><ol><li>Require review for VIP domains, confidence below 0.7, or a brand-new CRM record.</li><li>Give the reviewer three actions: approve, edit and approve, or reject with a reason.</li><li>Keep a short queue time, such as 30 minutes during business hours.</li><li>Log the input, the model output, the decision, the reviewer, and the time.</li><li>Each week, turn repeated rejections into a clearer rule so the same item stops coming back.</li></ol><h2>Common mistakes</h2><ul><li>Gating every step. The queue becomes the bottleneck.</li><li>Vague rules like “looks risky.” Write a condition you can test.</li><li>No record of who approved what.</li><li>No plan for after hours, so the queue grows overnight.</li><li>Reviewers rewriting from scratch instead of teaching the agent the fix.</li></ul><h2>Try it</h2><p>Open the <a href="https://gowithagentic.ai/tools/hitl-gates">human-in-the-loop gates</a> tool.</p>