What is human in the loop?
Sign off two hundred AI drafts a day and by the fortieth you are clicking approve without reading. The review exists; the oversight does not.
Human in the loop
HITL
Human in the loop describes an AI system that pauses and waits for a person to approve, edit or reject its work before the result takes effect. The person is part of the running system, not an auditor reading logs afterwards.
Two words carry the weight. In the loop means the work stops until somebody acts. Nothing has been sent, filed or charged yet.
That is different from a human on the loop, who watches a system that is already running and steps in when something looks wrong. Both are legitimate designs. Vendors use the first phrase to describe the second, so ask what happens when nobody clicks anything.
Put the checkpoint where the damage is
Review the step that is hard to undo, not the step that is easy to see. Drafting a summary is cheap to get wrong, because the next person notices. Sending it to a regulator is not.
So the checkpoint belongs immediately before the irreversible act: money leaving, a document filed, a message reaching a customer, a record changed in a system of truth. Everything upstream of that can run unattended, which is what makes the review affordable.
- 01Review what cannot be recalled, and let reversible steps run free.
- 02Show the reviewer the source next to the answer, or they are approving the layout.
- 03Count how often somebody rejects something. A queue that is never rejected is not a control.
- GatherUnattended. Cheap to redo.
- DraftUnattended. Nothing is final.
- PresentAnswer beside its evidence.
- DecideA person approves or rejects.
- CommitThe irreversible step, last.
Move the decision one station later and you are reviewing work that already happened. That is an audit, and it is a different product.
Related questions
01Does human in the loop mean the AI saves no time?
No, because reviewing is much faster than producing. A nurse confirming five ranked trial matches works in a fraction of the time it takes to search the registry unaided. The saving comes from the gathering and drafting, and you give a little of it back to keep the final call with a person.
02When can we remove the human review?
Once you can show, from your own logged decisions, that reviewers are rejecting almost nothing in a specific narrow case. Retire the checkpoint for that case only and keep it everywhere else. Removing it because the system feels reliable is how teams discover their reviewers had been quietly fixing things all along.
03Who should the reviewer be?
The reviewer should be someone who could have done the task themselves and carries the consequence of it being wrong. Oversight by a person with no domain knowledge produces approvals based on whether the output looks tidy. That is the most common way a compliant-looking process fails.
04Is human in the loop required by law?
Some automated decisions carry legal obligations toward human involvement, and which ones depends on your jurisdiction and the decision itself. Rules covering decisions with a major effect on a person tend to push toward meaningful review rather than a token approval screen. Get the specific obligation from your own counsel before designing the workflow around it.
