What is a human in the loop?
A human in the loop is a person positioned inside the decision path, whose approval is required before an outcome stands. The system prepares a recommendation and then waits. Nothing reaches the person affected until the reviewer has acted on it.
Also called: HITL
Why approval has to be possible as well as required
The arrangement is cheap to declare in a policy document and expensive to honour in a queue. A reviewer required to approve every outcome becomes the capacity limit of the whole deployment, and the pressure that follows is always towards approving faster. What the policy intended was a considered judgement, and what the deployment produces is a click. An inspector examining human oversight looks first at the median time each decision received, because that figure cannot be rehearsed.
- A required approval makes a named person answerable for the outcome.
- Review capacity becomes the throughput ceiling for everything behind it.
- Approval under volume pressure degrades into acknowledgement without examination.
- The reviewer needs evidence and the authority to disagree, or the control is decorative.
How a human in the loop works
- Decision categories requiring approval are defined before deployment.
- The system assembles a recommendation with its evidence and its confidence.
- Work pauses at the approval point rather than proceeding provisionally.
- The reviewer may amend or reject, with a reason required for any change.
- Identity and timestamp are recorded against the approval itself.
- Queue depth and review duration are monitored as oversight indicators.
How a human in the loop is evaluated
Oversight is assessed on whether the reviewer could realistically have decided otherwise. Measure the following per reviewer rather than in aggregate:
- median time spent on each decision awaiting approval
- rate at which recommendations are amended or rejected
- proportion of changes carrying a recorded reason
- queue depth against the review capacity available
- share of reserved categories that actually reach a person
Limitations, risks and common failure modes
- Approval rates close to total indicate a control that has stopped functioning.
- Reviewers see the recommendation before the evidence, which anchors the judgement.
- Oversight is assigned to staff junior to the person whose work they review.
- A backlog at the approval point is relieved by widening automatic approval.