Question 1

What is a human-in-the-loop approval flow?

Accepted Answer

An approval flow is the contract between an AI agent and the humans accountable for what it does. A human-in-the-loop (HITL) flow requires a person to approve each action before it executes &mdash; the agent drafts, the human signs off, then the action runs. Human-on-the-loop (HOTL) flows let the agent act autonomously while a human monitors and intervenes after the fact. Most production agents move from HITL to HOTL action-class by action-class as trust is earned with measurable evidence.

Question 2

When should an action require pre-approval vs post-review?

Accepted Answer

Pre-approval (HITL) for actions that are hard to reverse (outbound email to customers, public posts, payments, account changes), regulated (GDPR data exports, contract changes, anything legal team has flagged), or low-volume enough that a human can keep up. Post-review (HOTL) for actions that are reversible (internal notes, draft creation, status updates), high-volume enough that pre-approval would create a bottleneck, and auditable after the fact. Anything irreversible without strong reversibility (sending money) should stay HITL indefinitely unless the model is provably reliable on a 99.5%+ eval set.

Question 3

Which channel should approvals go through?

Accepted Answer

Match the channel to the team that owns the action. Slack works for fast-moving operational decisions (sales, support, marketing) where the approver lives in Slack all day. Email is right for slow, asynchronous, or compliance-heavy approvals where the audit trail matters more than turnaround time. Notion and HubSpot work when the approver already lives in those tools and you want context next to the approval. Custom web UI is right when the approval needs to show structured data (diff view, risk score, model reasoning) that does not fit a chat message.

Question 4

What does the audit trail need to capture?

Accepted Answer

Six fields, at minimum: agent_action_id, requested_at, approver_id, decision (approve/reject/edit), decision_at, and the model_input + model_output snapshot. For regulated workloads add: reasoning_trace (why the model proposed this action), data_source_ids (what the agent saw), and policy_match (which policy version was in effect). This is the line between an agent and a recorded business decision.

Question 5

What happens when an approval times out?

Accepted Answer

Three options, picked per action class. Auto-reject (safe default for high-risk actions) lets the action expire silently with a notification to the requester. Escalate (right for medium-risk, time-sensitive workflows) routes the pending approval to a second approver after the SLA. Auto-approve with audit flag (only for low-risk actions where blocking the workflow costs more than the occasional bad action) lets the action proceed but marks it for retroactive review. Write the timeout behavior into the blueprint &mdash; it is the most common gap in approval flow design.

Human-in-the-Loop Approval-Flow-Blueprint Human-in-the-Loop Approval Flow Blueprint

ErforderlichRequiredDie 5 PflichtfelderThe 5 required fields

Approver-RegelnApprover rules

Audit-TrailAudit trail

ComplianceCompliance

Fallback & Kanal-AusfallFallback & channel outage

Rollout: HITL → HOTLRollout: HITL → HOTL

Generierter BlueprintGenerated blueprint

Approval-Flow gebaut in 2 WochenApproval flow built in 2 weeks

Wie dieser Blueprint zu nutzen istHow to use this blueprint