Let AI move fast without losing human judgment on the big calls

Create a workflow 'AI action approval queue' that intercepts high-stakes AI-generated actions and requires human approval before execution. Goal: capture AI's speed while keeping human judgment on the consequential decisions. Configurable threshold for what triggers approval (ARR cutoff, segment size, action type, AI confidence).

Configure webhook trigger that other workflows call when they want to queue an action for approval. Payload schema: action_type (send_email / update_records / route_lead / make_decision), action_summary (one-line description), full_payload (the actual action data), requesting_workflow (originator), threshold_reason (why this needs approval — high-ARR / large-segment / low-confidence).

Multi-split by action_type: send_email → marketing lead approval; update_records → RevOps approval; route_lead → sales manager approval; make_decision → varies by decision type. Each routes to the right approver Slack channel. The approver matters — wrong approver leads to bottlenecks.

Configure Slack step that posts an interactive approval message to the right channel/user. Message includes: action summary, originator workflow, threshold reason, action preview (e.g. for email: subject + first 200 chars), and Approve/Reject buttons (or Slack thread-reply 'approve'/'reject'). Tag the specific approver.

Configure wait-until step that pauses workflow execution until approval response received. Timeout: 4 hours (configurable — high-ARR sends might warrant 1hr, routine updates might allow 24hrs). On timeout: default-reject (safer than default-approve) + post 'auto-rejected after timeout' notification.

Branch step on the approval decision: APPROVED → execute action (webhook back to originator workflow with 'go ahead'); REJECTED → archive with rejection reason + notify originator workflow to abort + post reasoning to #ai-governance for learning; TIMEOUT → same as rejected, with timeout reason logged.

Configure webhook call back to the originator workflow to proceed with the queued action. Include approval metadata (approver name, timestamp, any approver comments). Log to audit trail — every AI action that went through approval is fully traceable for compliance + later review.

On rejected branch: write the rejection to a structured learning log — which AI suggestions are humans rejecting, what's the rejection reason. Over time this reveals which AI behaviors need prompt tuning or guardrails. The reject log is the feedback loop that improves the underlying AI workflows.

Validate and publish. Monitor: approval volume per week, approval-rate (target: 80%+; below 60% suggests upstream AI workflows need tuning), median time-to-decision (target: under 30min during business hours), timeout rate (high = wrong approvers or low-engagement channel). The dashboard for responsible AI governance.