Definition
Human-in-the-loop means a person reviews, approves or corrects an AI system's work at defined points, especially where judgment, accountability or high impact is involved.
The design question is not whether to include people but where. Checkpoints belong on decisions that commit money, affect customers, carry compliance risk or depend on context a system cannot see.
Why it matters
Fully automatic is rarely the goal. The goal is that every decision is made by whoever, or whatever, is best placed to make it, and that someone is accountable. Human checkpoints are how a business keeps that accountability while still moving faster.
Placed well, they are cheap: a person approves a drafted refund in one click. Placed badly, they become a bottleneck, or a rubber stamp nobody reads. Good design shows the reviewer exactly what changed and why.
An example
An agent prepares supplier payments every Friday. It matches invoices to purchase orders and approves exact matches under a set amount on its own. Anything above the limit, or with a mismatch, goes to the finance manager with the difference highlighted, and their decision is logged against the payment.
Illustrative example, not a client project.
Common questions about Human-in-the-loop
Where should a human checkpoint go?
On steps that commit money, affect customers, carry legal or compliance risk, or depend on context the system cannot see. Low-risk, reversible steps usually do not need one.
Does human-in-the-loop slow everything down?
Not when it is designed around exceptions. Routine cases flow through automatically and people review only the cases that need judgment.
Related terms