Skip to content

AI

Human in the loop

A design where the system proposes and a person confirms, used wherever a wrong action is expensive to reverse.

Also called

  • HITL
  • Review queue

Human-in-the-loop is not a fallback for a system that is not good enough yet. It is a deliberate design for decisions where the cost of being wrong is asymmetric — approving a payment, denying a claim, sending a legal notice.

The design detail that matters is what the human is shown. A queue that presents a proposed value with no context turns a reviewer into a rubber stamp within a week. A queue that shows the source document with the relevant passage highlighted next to the proposal turns them into an actual check, and takes a fraction of the time re-keying would.

Confidence thresholds decide what reaches the queue. Set them from the evaluation data rather than intuition, and raise them as accuracy is demonstrated rather than as confidence grows.

It is what makes AI deployable in regulated or high-consequence work at all — and the queue design is what decides whether it stays a real control or decays into theatre.
Why it matters

Commonly misunderstood

What people get wrong

The claim

Human review means we haven't automated anything.

What is actually true

Confirming a highlighted value takes seconds; finding and keying it takes minutes. Most of the saving is in the search, not the keystroke.

Next step

Working through a human in the loop decision?

Tell us the situation. We will give you the tradeoffs as we see them, including when the answer is that you do not need what you are being sold.

No pitch deck. A 30-minute conversation about what you are trying to achieve.