human-in-the-loop (HITL)

basicwidespread

An operating mode in which the AI pauses before important steps and continues work only after human approval.

In plain terms

Imagine an intern who prepares a transfer or contract, but it only goes out after the boss has reviewed and signed it. Human-in-the-loop is the same in the AI world: the machine works, but at critical points a human decides whether to proceed.

This is necessary because the AI agent no longer just writes text but also acts: it sends emails, deletes files, transfers money, and places orders. A bad answer can usually be corrected before anyone uses it. A bad operation, however, can cause immediate damage and is often irreversible.

It differs from fully autonomous operation in that responsibility and the final say remain with the human. It differs from simple manual work in that the preparation, searching, and drafting are done by the machine, and you only need to pay attention at decision points.

How it works

System designers pre-identify risky steps: for example, payment, deletion, sending emails outward, or modifying a live system.
The agent works independently until it reaches such a step; there it stops and shows what it is about to do and why.
The human approves, rejects, or modifies the suggestion and can even provide new instructions.
The agent continues based on its decision, while the system logs who authorized what for later retrieval.
As trust grows, rules can be relaxed: routine steps run automatically, and only rare, high-stakes cases are escalated to humans.

An everyday example

You ask the AI assistant to book a flight to Barcelona. It searches for flights and fills in the data, but stops before payment and asks: 'Should I buy this ticket for 68,000 forints?' The money is only spent if you click to approve.

Who is affected

employeeIn an increasing number of roles, your job will be to review and approve work prepared by AI, meaning the responsibility remains with you. → Work & HR

entrepreneur, executiveYou must decide which processes to run fully automatically and where human approval is mandatory.

developerThe coding agent asks for your permission before running commands or modifying files, and you configure what it may do without asking. → Code

lawyer, citizenUnder the EU AI Act, high-risk AI systems must be designed so that humans can effectively supervise them. Such systems include AI used in credit assessment or recruitment. This does not necessarily mean approval before every single decision. → Law

A common misconception

MythIf a human approves the decision, the system is secure.

In realityApproval is only as valuable as the attention behind it. Anyone who clicks 'Allow' a hundred times a day without thinking is not verifying anything.

Limits and risks

Too many permission requests lead to approval fatigue: after a while, you click without reading, and protection becomes illusory.
People tend to trust the machine's suggestion even when it is wrong; this is called automation bias.
Human approval slows things down and costs money, creating a strong temptation to disable it precisely in inconvenient places.
You can only make good decisions if you understand what you are approving: if the system shows the planned step vaguely or incompletely, control is merely formal.

Similar, but not the same

human-on-the-loop – There, the human only monitors the self-running system and intervenes if needed; here, the step does not happen without approval.
autonomous agent – An autonomous agent carries out the task without human confirmation, whereas human-in-the-loop limits this independence at sensitive points.
human oversight – Human oversight is the broader legal and organizational principle, while human-in-the-loop is one specific implementation of it.

Related concepts

Broader

human oversight, AI safety, AI governance

Narrower

approval step, permission management

On the edge

AI-agent, autonomous agent, human-on-the-loop, safeguards, Automation bias

A brief history

The term originates from control theory and simulation, where the "loop" is the machine's feedback-operated cycle, which includes the human, such as the pilot in a flight simulator. Later, in machine learning, it was used for data checked and corrected by humans; with the spread of AI agents, it came to mean approval before action.

Where things stand

As of late 2026, the most widespread agent tools, especially coding agents, typically ask for permission by default before risky actions. The EU AI Act requires human oversight for high-risk systems, but the start of application of these rules may be postponed following an EU proposal. The debate is now about how much approval is reasonable before it turns into mere clicking.

Status: 2026 November

In the news

Appears in 3 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Articles urge risk-tiered rules as AI-generated code review burden growsAug 17, 2026 · Code
Claude Code's auto mode to become the default, with fewer human approvalsAug 10, 2026 · Code
Auto Mode to become the default in Claude CodeAug 3, 2026 · Code

Also known as

human-in-the-loop · human in the loop · HITL · ember a hurokban · ember a folyamatban · human approval

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.