Human in the loop
Place people at deliberate review points where uncertainty or consequence is high.
- 9 minutes
- intermediate
- Reviewed 2026-07-16
What is it?
Human-in-the-loop design creates explicit approval, review, correction, and escalation tasks inside an automated process.
Why does it matter?
A person can apply context that is unavailable to the system and can prevent an uncertain output from becoming an irreversible action.
The mental model
Automation handles the routine path; people resolve uncertainty, approve risk, and improve the system.
A simple example
A refund agent prepares the reason, amount, and evidence; a supervisor approves before the payment tool is enabled.
What it is not
It is not asking a person to rubber-stamp every step or using manual work to hide poor automation.
Learn this first
These ideas make the lesson easier to place.
- Agent loop
- Permissions
Your first 60 minutes
Use one focused hour to make the idea concrete.
- List decisions by consequence and uncertainty.
- Choose one approval threshold.
- Design the reviewer view with evidence, recommendation, and safe alternatives.
Build this first
Create an approval object that records who reviewed, what evidence they saw, their decision, and the final action.
When not to use it
Do not add approval to harmless high-volume steps when sampling, monitoring, or deterministic validation controls risk better.
What to learn next
Learn orchestration and auditability.