Human-in-the-loop is a system design where machine intelligence handles routine work but routes edge cases, low-confidence outputs, or high-stakes decisions to a human reviewer before action is taken. It's the difference between full automation and reliable automation.

A typical pattern is an AI agent or agentic workflow that processes inputs within defined boundaries, then pauses and escalates when a risk threshold is crossed. The routing logic is where the work actually lives: too aggressive and you ship failures; too conservative and you've built an expensive notification system that nobody trusts.

HITL is often a permanent architectural decision, not a temporary phase you graduate out of. Medical diagnosis, financial compliance, and code review keep humans in the loop because removing them increases liability — and because the cost of a wrong call is asymmetric. Over time, human decisions can feed back into the system to expand automation coverage, but the human doesn't fully leave. Guardrails define what the AI cannot do; HITL defines what it should not do alone. Both matter. Conflating them leads to systems that feel safe but aren't.