In plain terms
Picture a junior analyst who prepares the wire transfer but can't press Send; a manager signs off first. HITL builds that checkpoint into an AI workflow. The model does the heavy lifting, a human owns the decision.
Why it matters
It is the main route to using AI in high-stakes processes: you keep most of the speed, and accountability stays with a named person. It is also what regulators mean by human oversight; the EU AI Act requires it for high-risk systems. There is a cost: every checkpoint adds delay, and a reviewer approving hundreds of items a day drifts into rubber-stamping.
Example
An agent drafts refund decisions. Refunds under €50 go straight through; anything above waits in a queue for a support lead to approve, edit or reject. Rejections are logged and used to improve the prompt.
Most often confused with
HITL vs. Human-on-the-Loop (HOTL)
With HITL the system cannot take the next step until a person approves. With HOTL it proceeds on its own while a person monitors and can step in. The difference is the default: wait in HITL, continue in HOTL.
Under the hood
Implemented as an interrupt in the agent loop: before a tool call with side effects (payment, email, delete) the run is suspended and its state persisted, then resumed with the human's decision. Design choices: which actions are gated (by risk or confidence score), what the reviewer sees (diff, evidence, reasoning), what happens on timeout, and audit logging. Close relatives: human-on-the-loop (the system proceeds, a human monitors and can intervene) and human-out-of-the-loop (fully autonomous).