Glossary
Human-in-the-loop approval
Requiring a person to approve high-impact actions before automation carries them out, enforced in code the automation can’t bypass.
Also called Approval gate, Four-eyes principle.
In plain terms
An approval step only protects you if it’s enforced where the action happens, on the server, with a threshold the automation can’t change. A rule written in a prompt is a suggestion.
It also only works if the approver can see what they’re approving. A screen that hides what an action will really do turns approval into a formality.
See what happens without it
The failure pattern this protects against, privileged action without an enforceable approval gate, played out step by step.
Scroll the diagram sideways to see every system.
The limit was a suggestion, not a control.
What goes wrong:
- The refund limit lives only in the prompt: “never refund more than $500.”
- An email claims a manager has already approved an $8,000 refund.
- Persuaded, the agent calls the API, and no code checks the amount.
- The limit was a suggestion, not a control.
With the safeguard:
- The limit is enforced in the payment path, in code the model cannot change.
- The same email claims an $8,000 approval.
- The agent tries anyway. The gate holds the payment and asks a person.
- A person decides, and the agent cannot skip the check.
When it mattered
In 2020 Citibank meant to send an interest payment but wired nearly $900 million in principal to Revlon’s lenders. Three people approved it; the loan system’s screen made the real effect hard to see.