What Is Human-in-the-Loop? A Plain-English Definition
Human-in-the-loop (HITL) describes a workflow in which an AI system stops at set points and only continues once a person has made a decision. The idea matters most for AI agents, because they do more than write text: through tool use they take real actions in your business systems.
How does human-in-the-loop work?
The agent works on its own until it reaches an approval point. There it presents what it intends to do and waits. A person approves, edits or rejects the proposal, and only then does the agent carry out the action, such as sending a message or posting an entry.
Placement is what makes or breaks the design. Put a checkpoint where an action has effects outside the system or cannot easily be reversed:
- Before sending: a sales agent drafts a quote and a cover email; a person checks both before anything goes to the customer.
- Before paying or booking: an agent codes a supplier invoice to an account and cost centre; finance confirms before it is posted.
- Before changing master data: an agent proposes new payment terms for a supplier record; the owner of that data accepts them.
Research, summaries and drafts usually run without a pause. That keeps the agent useful while people decide only where it counts.
Agent frameworks support this directly. In LangGraph, interrupt() pauses the run, saves state through a checkpointer and continues when the graph is called again with Command(resume=...) (LangGraph docs). An approval can therefore wait for hours without the agent losing its place.
How does it relate to the EU AI Act?
Article 14 of the EU AI Act requires human oversight of high-risk AI systems. Human-in-the-loop is one way to provide that oversight, not the only one, and typical sales or service agents are usually not high-risk systems.
Article 14 says the people overseeing a system must be able to disregard, override or reverse its output and to halt it. Article 26 adds that deployers must assign oversight to people with the necessary competence, training and authority. Under the AI Omnibus adopted in 2026, these high-risk duties apply from 2 December 2027 for standalone systems and from 2 August 2028 for systems built into products (Addleshaw Goddard, as of October 2026). Treat this as orientation, not legal advice.
Related terms and common pitfalls
Human-in-the-loop names a specific approval step inside a workflow. Neighbouring terms work at a different level and are easy to mix up.
- Human oversight is the broader legal concept. It can be met without approving every action, for example through monitoring plus the ability to stop the system.
- Levels of autonomy describe how much an agent decides by itself overall. Human-in-the-loop marks where that autonomy ends.
- Guardrails are technical limits that apply without a person, such as a spending cap or blocked tools.
The main weakness is fatigue. Someone approving dozens of proposals a day soon stops reading them, which Article 14 names as automation bias. Keep approvals rare and meaningful: show what will change and why, not just an "OK" button. What sensible checkpoints look like in practice depends on the process, as the examples of AI agents for specific business functions show.
Frequently asked questions
Does the law require human-in-the-loop?
The EU AI Act requires human oversight for high-risk AI systems, but it does not prescribe a per-action approval step. For most everyday business agents, human-in-the-loop is a design choice rather than a legal duty. This is not legal advice.
What is the difference between human-in-the-loop and human-on-the-loop?
With human-in-the-loop, the system waits for approval. With human-on-the-loop, it keeps running on its own while a person monitors it and can step in.
Where should approval points go?
Before actions that leave your organisation or are hard to undo, such as sending, paying, booking or changing master data. Research and drafting can usually run without a pause.
Sources
- EU AI Act, Article 14: Human Oversight artificialintelligenceact.eu
- EU AI Act, Article 26: Obligations of Deployers of High-Risk AI Systems artificialintelligenceact.eu
- Addleshaw Goddard: EU AI Act – AI Omnibus formally adopted addleshawgoddard.com
- LangGraph documentation: Interrupts docs.langchain.com