# Human in the loop

> A person checks or approves an AI system’s consequential steps instead of letting every action happen automatically.

## What it means

Short, plain definition for AI at work, then the distinctions people search for: human in the loop (a person approves each consequential step), human on the loop (a person supervises and can step in) and human out of the loop. Where the phrase comes from (control systems and simulation, then labelling and review in machine learning) — stanford.edu/ai-definitions/what-is-human-in-the-loop) before quoting. Examples: approving an AI-drafted reply; reviewing rows an assistant drafted before they are added.

## A practical example

An assistant drafts twelve rows for a project table. A person checks the dates and owners, removes two weak rows, then approves the remaining ten.

## Put it into practice

- Name the concrete situation before applying the label. Human in the loop is useful only when it changes a choice, boundary or next action.
- Make the state visible: write the owner, time, source or stopping condition down instead of relying on memory.
- Review what happened. Keep the part that reduced ambiguity and change the part that added ceremony without helping the work.

## How it relates to Plexii

Plexii proposes changes as cards you approve before anything changes (/help/ai-assistant); a Desk agent’s results arrive as Proposed changes cards unless you tick Write without asking me first (/help/widget-desk-agent) — state both honestly. Keep it 300–400 words and send readers to /blog/approve-ai-changes for the practical guide rather than repeating it.

## Sources

- [Stanford HAI: What is human-in-the-loop AI?](https://hai.stanford.edu/ai-definitions/what-is-human-in-the-loop)