What is human-in-the-loop (HITL)?

Length:

6 min

Published:

September 9, 2026

What is human-in-the-loop?

Human-in-the-loop (HITL) is a design in which an automated system stops and waits for a person before it acts. The person is not observing from the outside. They are a step inside the loop, and nothing downstream runs until they approve, edit or reject what the machine produced.

Machine learning uses the phrase in a second, older sense: people supply the judgement a model lacks by labelling training data or correcting outputs that then feed back into training. Both senses share one property. Take the human out and the loop still closes, only now without the check.

So the useful question is never whether a human is involved. It is where they stand, and whether their approval carries any weight.

In plain words

Think of a bank transfer that needs a second signature. One person prepares it, the money does not move, and somebody else has to look at it and sign. Human-in-the-loop is that second signature, wired into software. Which raises the same question banks had to answer: does the second signer actually read the transfer, or sign fifty of them before lunch without looking?

In the loop, on the loop, out of the loop

The three phrases sound interchangeable and describe different systems.

In the loop. The system pauses and waits. Nothing happens without approval. This is the slowest arrangement and the only one where a human decision comes before the effect.
On the loop. The system acts and reports. A person monitors it and can intervene, stop it or reverse what it did. Faster, and it assumes the action can be undone.
Out of the loop. The system acts alone. Any oversight happens afterwards, through logs, audits and spot checks.

Treat the choice as a dial rather than a switch. Two questions set it: how bad is a wrong action, and can you undo it? An agent that drafts text can sit on the loop. An agent that issues refunds or deletes production data belongs in it.

What the law actually requires

This is worth getting precise, because "we keep a human in the loop" often gets said as a synonym for "we are compliant". Article 14 of the EU AI Act covers human oversight of high-risk AI systems. It requires them to be built so that people can oversee them effectively. An overseer has to understand the system's capacities and limitations, interpret its output, and decide in a given situation not to use it. They must also be able to "disregard, override or reverse the output" and to interrupt the system through a stop button or similar procedure.

Read those verbs again. Override, reverse, interrupt. That is oversight of a running system, not mandatory approval before each action. The Act demands the capability to intervene rather than a signature on every step. One case is stricter and named explicitly: for biometric identification, no decision may be acted on unless the identification has been "separately verified and confirmed by at least two natural persons".

The same article also names the failure mode. Whoever deploys the system has to stay aware of "the possible tendency of automatically relying or over-relying on the output produced by a high-risk AI system (automation bias)". The law anticipates that the human in the loop stops reading.

How it works in an agent

Three mechanisms cover most of what teams build.

Pause and resume. LangGraph exposes an interrupt() call inside a node. When it fires, the framework "saves the graph state using its persistence layer and waits indefinitely until you resume execution", so a pause costs no running process. Its documented patterns are approval workflows, review and edit state, interrupts in tools, and validating human input.
Protocol-level questions. MCP elicitation lets a tool server ask the user something mid-run instead of guessing or failing.
Permission prompts. Coding agents ask before they write a file or run a command, which is the same pattern at the smallest possible grain.

The mechanism is the easy half. A pause only works if answering it is cheap and the request carries enough context to decide on. A prompt that says "allow this action?" with no diff and no consequence trains people to click yes.

What to watch out for

A gate placed after the work is not a gate

We learned this on our own content pipeline. An agent researched a topic, wrote the article and opened a pull request, and a human reviewed it at the end. Every individual step worked. The queue did not. When we last counted, nine finished bilingual pairs, eighteen pages of the site, sat unmerged, the oldest of them for 86 days. Production had outrun review by two months.

The human was in the loop the whole time. They were just standing after the expensive part, where the only options left were approve, reject or rewrite someone else's finished draft. In September 2026 we moved the gate to the front, so a person now picks the topic before the agent invests any research. Same number of humans, far less waste, and a "no" that costs a sentence instead of a week.

If the only human checkpoint sits at the end of the work, it is not a control. It is a queue.

The rest of the failure modes

Approval without attention. A reviewer who sees forty requests an hour approves them at a rate unrelated to reading them. Measure how long a decision takes, not how many get made.
Skill fades in the monitor's seat. Lisanne Bainbridge argued this in Ironies of Automation (Automatica, 1983): automate most of a task and the operator loses the practice needed for the rare intervention that matters. She wrote it about industrial control rooms. It transfers to anyone reviewing agent output.
Diffused accountability. When a person approves what a model proposed, ownership of the outcome gets blurry. Write down who answers for it before you ship the workflow.
A pause that hides its cost. Waiting on a human is often the slowest part of an agent run. Batch approvals or narrow what needs one, rather than sprinkling prompts everywhere.

What is MCP elicitation? - The protocol-level way an AI tool asks a person a question in the middle of a task.
What are AI guardrails? - The automated half of the same problem: constraints that hold when no human is watching.
What is AI governance? - Where human oversight sits inside the wider set of rules, roles and accountability.
What's an agent? - What it is you are putting a human in the loop of.
What are agentic workflows? - How multi-step agent runs are structured, and where a pause can go.

Back to insights

Want to stay one step ahead?

Don't miss our best insights. No spam, just practical analyses, invitations to exclusive events, and podcast summaries delivered straight to your inbox.