Human-in-the-loop: staying in control of what AI does
Human-in-the-loop describes an AI system built so that a person can review, approve, correct or stop its important decisions before they take effect. It is the main safety net for agents that can act on your behalf.
Data as of October 22, 2026
Key figures
| Level | Intermediate |
|---|
The principle
An AI agent can chain actions together: search, fill in a form, send a message. Human-in-the-loop means building in checkpoints where the agent pauses, shows what it is about to do and waits for an answer: approve, edit or reject. The LangGraph framework documentation describes its “interrupts” this way: pausing execution at specific points to wait for outside input, for example before an API call, a database change or a financial transaction.
A concrete example
You ask an assistant to reply to quote requests that arrive by email.
- The agent reads the messages, drafts a reply and works out a price from your rate card.
- Before sending, it shows you the draft and the amount, then waits for your go-ahead.
- You fix a discount that was applied wrongly, then approve. The email goes out.
Without that checkpoint, the pricing error would have landed straight in the customer’s inbox.
Why it matters
Models make mistakes, and an agent can be manipulated by booby-trapped text hidden in a web page or document; this is called prompt injection. Human approval limits the damage when it happens. It is also a regulatory topic: Article 14 of the EU AI Act (official text) requires high-risk systems to be overseeable by people who, among other things, stay aware of the tendency to over-rely on the machine and can stop the system. To decide how much oversight a given task needs, use our Can I trust an agent with this? tool.
Common mistakes and myths
- “A human in the loop makes the system safe.” Only if that person has the information and the time to judge. Rubber-stamping out of habit protects nothing.
- “Everything must be approved.” Too many confirmation prompts cause fatigue and lead people to click without reading. Focus on the risky actions.
- “It replaces other protections.” It complements technical guardrails and sandboxing, which limit what an agent can do even when nobody is watching.
Related terms: agentic workflow, AI agent, guardrails, prompt injection, sandboxing.
Related terms
- AI agent · An AI that does more than answer: it plans and carries out a series of actions with tools to reach a goal you set.
- Guardrails · Checks placed around an AI model to filter what goes in and what comes out, and to block unwanted answers or actions.
- Agentic workflow · A sequence of steps in which an AI model plans, uses tools, checks and revises its work instead of answering in a single pass.
- EU AI Act · The European law that regulates AI systems according to their level of risk, with bans and transparency duties.
- Prompt injection · An attack that slips hidden instructions into text an AI reads, to make it ignore its original instructions or act against its user.
- Sandboxing · Running an AI agent inside an isolated environment with restricted files, network and commands, so any damage stays contained.
Try it
Full glossary
Frequently asked questions
When should a human approve an AI agent's step?
Mainly before actions that are hard to undo: payments, messages sent outside the company, deleting data, publishing. Reading and drafting tasks can often run without approval at every step.
Is human-in-the-loop a legal requirement?
For AI systems classified as “high-risk”, Article 14 of the EU AI Act requires that they can be effectively overseen by people. For personal use it is good practice rather than an obligation. This is general information, not legal advice.
Doesn't approving every step slow everything down?
Yes, that is the trade-off. The goal is to put approvals where a mistake would be costly and let the agent move on its own elsewhere.
Is clicking “Approve” enough to keep things safe?
No. If the person approves on autopilot, the check becomes a formality. They need to see clearly what is about to happen and have time to verify it.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.