Learn AI /AI Agents and Automation /Human in the loop
Human in the loopHITL
- Starter
- AI Agents and Automation
- Policy officers
- Managers & front offices
- Human rights practitioners
Also called: HITL, human-in-the-loop.
Back to the milestone you came from
In five seconds
What it is
A system setup where a person reviews, guides or approves an AI-driven step before the workflow continues or completes.
In plain words
A person is kept in the process so the system does not act completely on its own.
An everyday example
A traveller in Bogotá wants to cross the city during heavy rain. A route-planning app suggests three options using live traffic data, but before booking a ride it asks the traveller to confirm the safest pick-up point and whether they are comfortable with a higher toll road fee. The traveller checks the street names, rejects one option near flooding, and approves another. Only then does the app complete the booking.
A travel app example makes the human role look simple and personal. In real AI systems, the human may be a caseworker, analyst or supervisor following formal rules, and there may be many review points, audit logs and multiple people involved. Also, some HITL systems use the human after an action for correction or escalation, not only before action.
Remember thisHuman in the loop means a person has a real decision point inside the automated workflow, with enough information and authority to approve, correct, reject or stop the proposed action.
Understand it
Understand it
Why it matters
For non-specialists, the key question is not whether AI is used, but where responsibility sits when something important happens. Human in the loop helps organisations control risk by putting a person at the points where errors could cause harm, unfair treatment, reputational damage or unsafe action. It can improve quality and trust, but only if the human has enough information, time, authority and training to make a real check. If the review is rushed or purely symbolic, HITL can give false reassurance rather than real control.
How human in the loop works
Show where a person can pause, check and steer an AI-driven workflow before the final action or output.
01
1. Request starts
A user request, task or new case enters the workflow.
Draw a rounded box labelled 'Request' with small icons for text and data entering from the left.
02
2. System prepares a draft or action
The AI system analyses the input and proposes a response, classification, plan or tool action.
Draw a second box labelled 'AI draft / proposed action' connected by an arrow from step 1.
03
3. Human review gate
A person checks the proposal, especially for accuracy, risk, fairness, safety or policy compliance.
Draw a highlighted decision box labelled 'Human review' with a person icon. Use a contrasting border to mark it as the control point.
04
4. Human choice
The reviewer can approve, edit, reject, or send the case for escalation.
From the review box, show four short labelled branches: 'approve', 'edit', 'reject', 'escalate'. Merge approve and edit into the next step; reject and escalate end in stop or specialist review.
05
5. Final output or action
Only the approved or corrected result is sent, saved or carried out.
Draw a final box labelled 'Final output / action' for approved flows, plus a side box labelled 'Stopped or escalated' for the other branches.
A left-to-right step diagram with five boxes connected by arrows. Read from request to final result. The key idea is that the process reaches a review gate where a person can approve, edit, reject or escalate before the system continues.
On hover or tap, each step reveals a short note with an example such as message drafting, document tagging or booking approval.
Describe this without the diagram
The diagram shows a five-step process. First, a request enters the system. Second, the AI system produces a draft or proposed action. Third, the workflow reaches a human review gate, which is the main point of the diagram. Fourth, the human can choose one of four paths: approve, edit, reject or escalate. Approve and edit both lead to the final output or action. Reject and escalate do not continue normally; they stop the process or send it to a specialist. The purpose is to show that the human is placed inside the workflow, not outside it.
With reduced motion: A static five-box sequence is shown with labelled arrows. The human review box is highlighted, and the four possible outcomes from review are visible at once: approve, edit, reject or escalate.
Human in the loop means a person is deliberately placed inside an automated process so the system does not run entirely on its own. The person may check an input, approve a step, correct an output, or stop the process.
What goes in can be a request, some data, and a rule about when a human must step in. For example, a system may draft a reply, sort applications, or prepare a route plan. A person then reviews the draft or decision point before the system continues.
What happens next depends on the design. In some systems, the human must approve every action. In others, the system acts on low-risk tasks by itself and only asks for help on unclear, risky or unusual cases. That human input can also become feedback for improving later behaviour, though improvement is not automatic.
What comes out may be an approved message, a corrected classification, a blocked action, or an automatically handled low-risk case. This differs from full automation because the workflow has defined points for human intervention, and from general human oversight, which may operate outside the live workflow.
How “Journey planning with a travel assistant” maps onto the real thing
| In the story | In the AI system |
|---|---|
| The app suggesting several routes | The system generates options from available data |
| The traveller checking safety and cost | A human reviews risk, context or preferences the system may miss |
| Approval before booking | A gated action that needs human confirmation before the system proceeds |
A worked example
A programme officer in a country office is preparing a progress report for first reading. An automated workflow drafts paragraphs from implementing partner updates, but cannot place them in the file without the officer's approval.
Starting from: Partner update: “Training has started. Attendance records are still being verified.”
- The system proposes: “The planned training was completed and attendance was verified.”
- Because the completion claim is not supported by the source, the workflow routes the paragraph to the programme officer rather than inserting it automatically.
- The officer changes the text to: “Training has started, and attendance records are being verified.”
- After approval, the workflow places the corrected paragraph in the draft progress report and records the review.
Result: The workflow saves drafting time without allowing an unsupported claim into the file. The human intervention is part of the execution path, not a review conducted after circulation.
In a United Nations workplace
During the annual workplan review, a training officer in a procurement unit uses an automated workflow to turn cleared workplan actions into draft course entries. Before the workflow updates the learning portal or circulates invitations, it pauses for the officer to check the dates, intended audience and source text, then approve or correct each entry.
The approval gate sits before the system takes an external action. The training officer can correct the proposal, reject it or send an exception back for follow-up, while the workflow records the decision and continues only after approval.
Before you try thisBid information, vendor evaluations, personal data and other procurement-sensitive material must not be entered into a tool that has not been approved for it. The review must also be substantive: an approval button does not provide meaningful oversight if the reviewer cannot inspect the source, correct the output or stop the action.
Where you may hear it
- A vendor demo in a procurement meeting
- An automation proposal before an ICT governance board
- A front-office clearance discussion before automated circulation
- An internal audit walkthrough of approval gates and exception handling
Avoid confusion
Avoid confusion
Often confused with
Human oversight
Human oversight is the processes that let people monitor AI use, review its outputs and effects, and intervene when needed.
Meaningful human control
People with real authority must have enough information and practical ability to shape, limit, change or stop AI-supported actions, especially in high-risk situations.
Copilot
A tool that helps a person complete digital tasks by drafting, searching, and sometimes taking limited actions under human direction and review.
Common misconceptions
Common misconception: If a human is in the loop, the system is automatically safe or fair.
Correction: Not necessarily. Safety depends on where the person is placed, what they can see, whether they can override the system, and how much time they have. A nominal review step can miss serious problems if the workflow pushes people to approve quickly.
Common misconception: Human in the loop means a person checks every single output.
Correction: Some systems work that way, but many only send selected cases to a person, such as low-confidence, high-risk or unusual cases. The term describes a designed human checkpoint inside the process, not one fixed review pattern.
Common misconception: It is the same as human oversight.
Correction: They are related but not identical. Human oversight is broader and can include monitoring, governance, audits and after-the-fact review. Human in the loop is narrower: the person directly participates in the live workflow before it completes or proceeds.
Quick check
Which example best matches human in the loop?
The defining feature is that a person is part of the live workflow before the action is completed. A monthly report is oversight, not HITL inside the specific task. Full auto-send has no human in the loop.
Go deeper
Go deeper
The simple version
A person is kept in the process so the system does not act completely on its own.
The technical version
HITL is a workflow design pattern in which one or more human interventions are built into the execution path of an automated system. The intervention may occur before an action, at a decision threshold, after a model proposes an output, or during exception handling. Common mechanisms include approval gates, confidence-based routing, sampling for review, escalation rules and correction interfaces. In agentic systems, the checkpoint may sit before tool use, before sending a message, before updating a record, or before taking a real-world action through an API or connector. HITL is often confused with broader governance terms, but the distinctive feature is operational participation in the live process rather than only external supervision.
Reviewed external resources
Primary sourceModel Context Protocol (opens in a new tab)Model Context Protocol
This link explains the Model Context Protocol and suits technical readers who want to understand how systems connect to tools and data sources that may need review gates.
Deeperlink checked 2026-07-31
EncyclopaediaHuman-in-the-loop (opens in a new tab)Wikipedia
This link gives a broad introductory overview of the term and suits readers who want a quick starting definition.
Intermediatelink checked 2026-07-31
ArticleBuilding effective agents (opens in a new tab)Anthropic
This link explains how agents operate in practice and suits readers who want to see where human checkpoints may fit in larger agent workflows.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Human in the loop (opens in a new tab)YouTube search
This search results page offers multiple video explanations of human in the loop and suits readers who prefer visual or spoken introductions; results quality will vary.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Human in the loop (opens in a new tab)Reddit search
This search results page shows community discussions and examples from different users, which may help readers see common questions, but it suits cautious readers because quality and accuracy will vary.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Human in the loop means a person has a real decision point inside the automated workflow, with enough information and authority to approve, correct, reject or stop the proposed action.