---
title: "Human in the loop: definition | KYA glossary"
url: https://knowyouragent.ch/en/knowledge/glossary/human-in-the-loop
lang: en
updated: 2026-10-12
license: © Innopulse Consulting GmbH
---

# Human in the loop

Human in the loop describes an approach in which a person reviews certain actions of an AI agent before execution and approves or rejects them.

Not every action of an agent should happen automatically. For high amounts, new counterparties or unusual patterns, the agent stops and presents the action to a person. Only after that person's approval is it executed. Related are “human on the loop”, in which a person only monitors and can intervene, and “human out of the loop” for fully automatic flows.

Human in the loop is effective when thresholds are clear, the approval is recorded and people are not flooded with so many requests that they just confirm. The best complement is a mandate with limits: whatever lies within it runs automatically, whatever goes beyond it goes to the person.

## Example

A purchasing agent orders supplies up to CHF 2,000 itself. It presents an order of CHF 7,500 to the head of purchasing; she approves it, and a signed decision records who approved it and when.

## How it differs

Human in the loop decides case by case; a mandate sets the limits in advance. Approval without a mandate scales poorly, and a mandate without approval for exceptions is rigid. Together they cover routine and exception.

## In practice

- Set thresholds above which a person decides, instead of presenting every action.
- Record every approval as a signed decision with person, time and reason.
- Track the number of approvals; if it rises sharply, the mandate or thresholds are set incorrectly.

## Role in KYA

KYA checks whether an action lay within the mandate and whether a decision exists. A human approval becomes verifiable as a signed decision; without this evidence, it remains open whether the exception was really approved.

Related check: [Scope of the action](https://knowyouragent.ch/en/method/checks) (`intent_scope`). Is the action within the mandate's action type, amount limit, currency and counterparty?

[Verify now: open a matching example](https://knowyouragent.ch/en/verify?beispiel=abgelehnt)

## Frequently asked questions

### What does human in the loop mean for AI agents?

A person reviews certain actions of an agent before execution and approves or rejects them, typically for high amounts or unusual cases.

### What is the difference from human on the loop?

With human in the loop, the agent waits for approval. With human on the loop, it acts on its own, and a person monitors and can intervene.

### Is human approval enough as a safeguard?

Only if it is used deliberately. If people are flooded with requests, they eventually confirm without checking. Clear thresholds and a mandate for routine cases are therefore important.

### How does an approval become provable?

As a signed decision naming the action, the approving person and the time. This makes it possible to show later that an exception was approved.

### How does KYA help?

KYA checks whether an action lay within the mandate or was covered by a decision. If the decision for an exception is missing, the certificate shows it.

## Sources

- [NIST: Least Privilege](https://csrc.nist.gov/glossary/term/least_privilege)

## Related terms

- [Agent mandate](https://knowyouragent.ch/en/knowledge/glossary/agent-mandate)
- [Decision token](https://knowyouragent.ch/en/knowledge/glossary/decision-token)
- [Policy decision point (PDP)](https://knowyouragent.ch/en/knowledge/glossary/policy-decision-point)
- [Least privilege](https://knowyouragent.ch/en/knowledge/glossary/least-privilege)

---
Quelle: https://knowyouragent.ch/en/knowledge/glossary/human-in-the-loop · KYA – Know Your Agent