---
id: con-human-in-the-loop
schema: concept/1
layer: garden
status: established
name: Human in the loop
aka: ["HITL", "approval steps", "human review", "confirmation before acting", "sign-off"]
summary: >
  Designing AI work so a person reviews or approves at the points that matter, especially before
  anything is sent, published, paid or deleted.
first_seen: unknown
relations:
  builds_on: [con-steering]
  evolved_from: []
  related: [con-automation, con-privacy-and-data]
  parallels:
    - idea: "sign-off authority"
      field: "finance"
      note: "Spending above a threshold needs a named person's approval, however routine the request."
    - idea: "a pilot monitoring the autopilot"
      field: "aviation"
      note: "The system flies; the pilot watches, and takes over when something looks wrong."
implementations:
  - { provider: openai, name: "ChatGPT agent confirmations before actions" }
  - { provider: anthropic, name: "Claude permission prompts" }
  - { provider: peter-yang, name: "/human-review skill: edit and comment on AI output in the browser, then send all feedback at once", resource: res-other-peter-yang-human-review-skill }
competencies: [cmp-steering, cmp-use-responsibly]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "Turns 'know when a person has to decide' into a concrete design choice for agents and automations."
---

The more AI can do on its own, the more it matters where a person steps in. Human in the loop means deciding, on purpose, which steps a person must review or approve. The aim is a few well-placed approvals.

A useful way to choose: look for steps that are hard to undo or that affect other people. Sending an email to a client, publishing a post, paying an invoice, deleting files and making decisions about someone's job, money or health all deserve a human check. Drafting, sorting and summarising usually don't need approval each time, though they need occasional spot checks.

Approval only works if the person reads what they approve. Rubber-stamping fifty AI drafts a day without reading them is a loop in name only. Keep approval steps few enough that they stay meaningful.

## Example

An agent prepares weekly reminder emails to families about overdue library books. It drafts and queues them, but a staff member reviews the list each Thursday before sending, because one wrong email to the wrong family would damage trust.

## How it connects

- A structured form of **steering**.
- The safety design for **agents** and **automation**.
- Required wherever **privacy and data** or people's welfare is at stake.

## Try it

Pick a task you'd like AI to do for you regularly. Write down each step, then circle the one you'd never want done without you seeing it first. That's your approval point.