A workflow where an AI agent performs tasks but requires human approval at critical decision points, before the action executes. Common in high-stakes domains: a legal agent drafts a redline but a lawyer approves it; a cybersecurity agent recommends containment but an analyst clicks 'execute.' HITL is the strictest of three oversight postures — a human *in* the loop approves each action, a human *on* the loop monitors an autonomous agent and can intervene, and a human *out of* the loop only audits after the fact. HITL balances automation speed with human judgment, and is required for high-risk systems under EU AI Act Article 14.
Example
An AP automation agent processes 400 invoices per week. The agent extracts data, matches POs, and categorizes—but routes any invoice over $10K or with discrepancies to the AP manager for explicit approval via an Approval Gate. 85% of invoices flow straight through; 15% need human review—but the human review takes 30 seconds (confirm/reject) instead of 5 minutes of data entry.
Frequently asked questions
When should an AI agent require human approval?
Require approval when: (1) the action is irreversible or expensive (sending money, deleting data, publishing public content), (2) the agent's confidence is below threshold, (3) regulations require human accountability (clinical decisions, legal advice, certain HR decisions), or (4) the cost of an error exceeds the cost of human review time. As confidence builds and patterns become well-understood, lower-risk human gates can be removed. Treat HITL gates as a knob, not a constant.
What's the difference between human-in-the-loop and human-on-the-loop?
In-the-loop means the agent pauses and a human approves each significant action before it runs. On-the-loop means the agent runs autonomously while a human monitors from outside and can intervene, halt, or roll back. In-the-loop trades throughput for control and fits high-consequence actions; on-the-loop trades some control for speed and fits proven, reversible workflows. Most production agents use both, assigned per action rather than per agent.