What is human-in-the-loop? Explained with a manager signing cheques

In one sentence: human-in-the-loop means an AI system prepares the work, but a named person checks and approves the important steps before they happen, and can change, reject or stop them.

The everyday picture

Think of a busy office where a clerk prepares the cheques. The clerk fills in the payee, the amount and the date. That saves the manager hours each week.

But no cheque leaves the building until the manager signs it. The manager glances at each one. Most get signed. Some get sent back with a note: wrong amount, wrong supplier, ask me first. Every signed cheque is written in the logbook.

Human-in-the-loop works the same way. The AI is the clerk. A person is the manager. The signature is the moment the AI must wait for a yes.

How it works, step by step

1. The AI prepares an action (the clerk fills in the cheque). An AI model or AI agent drafts something: a reply to a customer, a refund, a change to a record.
2. The system pauses (the cheque goes on the manager's desk). For steps the business has marked as needing approval, the software stops and waits. In OpenAI's Agents SDK, for example, a tool marked as needing approval pauses the run until someone decides.
3. A person reviews it (the manager reads the cheque). The reviewer sees what the AI wants to do and why, then approves it, edits it or rejects it.
4. The approved action runs (the cheque is sent). Only after a yes does the email go out or the payment move. A rejected action goes back to the AI with a note, so it can try another way or stop.
5. The decision is recorded (the logbook). A good setup keeps an audit trail: what the AI proposed, who decided, and when. That record is how you check the process later.

Where the picture stops

• The clerk does not learn from the note on its own. A human clerk remembers being corrected. Most AI systems do not, unless someone updates their instructions or data.
• Not every cheque needs a signature. In real systems you choose which actions pause. Low-risk steps, like reading a document, often run without asking, or nobody would get any work done. Tools such as Claude Code let you choose how often the AI stops to ask.
• A tired manager signs anything. People tend to trust a machine's suggestion too much. The EU AI Act calls this "automation bias" and asks that overseers stay aware of it. A signature only helps if the reviewer really reads.

Why it matters for your business

• Lower risk. A person catches a wrong refund or a strange email before a customer sees it, not after.
• Speed where it is safe. The AI does the slow preparation. People spend their time only on the decisions that matter.
• Rules may require it. For high-risk AI systems, Article 14 of the EU AI Act requires that people can understand, override and stop the system, for example with a "stop" button.
• Clear responsibility. NIST's AI Risk Management Framework says human and AI roles should be clearly defined. Knowing who signs makes it clear who answers for a decision.

An example

Imagine a 15-person online furniture shop that gets dozens of refund requests a week.

They set up an AI assistant that reads each request, checks the order and drafts a reply. Replies that only answer a question go straight out. Any refund over a set amount pauses and lands in the support lead's queue. She approves most with one click, fixes a few, and rejects one where the customer is asking about the wrong order. Each decision is saved with her name and the time.

Frequently asked questions

What does human-in-the-loop mean in AI?

It means a person is part of the process and must approve, correct or reject certain AI outputs or actions before they take effect.

What is the difference between human-in-the-loop and human-on-the-loop?

In the loop, the AI waits for a person to approve each important step. On the loop, the AI acts on its own while a person watches and can step in or stop it.

Does human-in-the-loop slow everything down?

Only for the steps you choose to pause. Most teams require approval for actions that move money, change records or reach customers, and let low-risk steps run on their own.

Is human review required by law?

For some uses, yes. The EU AI Act requires human oversight for high-risk AI systems. Other rules depend on your country and industry, so check the ones that apply to you.

Related explainers

• What is an AI agent? Explained with an assistant you send on errands
• What is an LLM? Explained with a new hire who read the whole library
• What is MCP? Explained with an office of service counters

The DevDrift builds AI automation with approvals and audit trails: see how.

Sources

• European Commission AI Act Service Desk: Article 14, Human oversight
• NIST AI RMF, Appendix C: AI risk management and human-AI interaction
• OpenAI Agents SDK: Human-in-the-loop
• Claude Code docs: Choose a permission mode