Glossary
Human in the loop
A person who can approve, edit, or stop an agent action before that action counts.
Human in the loop means a person stays in the run. The agent can draft, look things up, and propose a next step. A person approves, edits, or stops the step that would send, pay, delete, or otherwise leave the organization.
OpenAI’s business guide, read November 7, 2026, says guardrails should seek confirmation before many real-world actions and ask for oversight on sensitive tasks such as sending email. ChatGPT’s Work guide, read the same day, tells you to say when the task should stop for review, and to follow the notice if a task pauses for a safety check. A guardrail is the check inside the harness. An approval gate is one configured stop, such as “ask before this write.” Autonomy is how many steps are allowed before that person is asked.
The workplace guide is What is an AI employee?. The click-by-click case is Can an AI agent use your computer?.