# Human in the loop vs human on the loop: choosing the control model

> Human in the loop vs human on the loop: a person confirms before the action, or supervises after it starts. How to choose by consequence, with a stop path.

- URL: https://www.ornsoft.com/blog/human-in-the-loop-vs-human-on-the-loop/
- Published: November 13, 2026
- Updated: November 13, 2026
- Author: OrNsoft team, OrNsoft Corporation
- Category: Trusted AI
- Publisher: OrNsoft (https://www.ornsoft.com/)

Human in the loop means a person confirms an AI action before it happens. Human on the loop means the action happens and a person supervises, with the power to stop or reverse it. The choice between them is not about how much you trust the model. It follows the consequence of the action: what it touches, whether it can be undone, and who it reaches.

**Human in the loop puts a person before the action. Human on the loop puts a person over the action, with a way to stop it and a way to undo it.**

This article is for operators and technical owners who are deciding how much control an AI system needs before go-live. It compares the two models, gives a way to choose, and says what each one requires to work. It builds on [approval gates for sensitive automated actions](/blog/approval-gates-for-ai-actions/), which covers how to design the gate itself.

> **Key takeaways**
>
> - **In the loop is for irreversible or external actions.** The action waits for a named approver.
> - **On the loop is for reversible actions inside your operation.** It can run at machine speed under supervision.
> - **Supervision needs tools.** Alerts, a view of the work, a stop path and a way to undo.
> - **One system can use both.** Decide per action, not per system.
> - **Loosen with evidence.** Move an action to a lighter model only when the record supports it.

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

The difference is timing and power. In the loop, the person acts before the system does: they approve, edit or refuse. On the loop, the system acts and the person watches, ready to intervene.

![A table comparing human in the loop with human on the loop on timing, what the person does, what each suits, what each needs and speed.](../../../assets/blog/human-in-the-loop-vs-human-on-the-loop/loop-comparison.png)

Neither model is more responsible than the other in the abstract. A confirmation step that shows a person only a green button invites automatic approval. A supervision role with no alerts and no stop control is a role in name only. Each works when it is built to work.

## How do you choose between them?

Start from the action. Two questions do most of the work.

![A decision flow: if an action cannot be undone, use in the loop; if it can, and it leaves your organization, use in the loop; otherwise use on the loop with a stop path.](../../../assets/blog/human-in-the-loop-vs-human-on-the-loop/control-model-flow.png)

1. **Can the action be undone?** If not, a person confirms before it runs.
2. **Does it leave your organization?** A message to a client, a payment or a filing cannot be pulled back by changing an internal record. These also wait for a person.

Actions that can be undone and stay inside your operation, such as drafting a reply, updating an internal field or creating a task, can run on the loop, with sampled review after the fact. Your owners make the final list. Re-check an action's tier whenever it changes what it can touch.

## What does in the loop require?

The control only works if the person can decide in seconds and the process cannot stall.

- **A named approver per action type, with a backup.** A gate that waits for "someone in finance" waits forever.
- **Context.** The proposed action exactly as it would run, the inputs behind it, and why it was flagged.
- **Three choices.** Approve, edit or refuse.
- **A timeout rule.** If nobody answers, the action does not run, the owner is alerted, and the case goes back to a queue a person already works.

NIST's [Generative AI Profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) names automation bias, excessive deference to automated systems, as a risk, which is why the approver should see the inputs and not only the output. The design is set out in [approval gates for sensitive automated actions](/blog/approval-gates-for-ai-actions/).

## What does on the loop require?

Supervision is a job with requirements.

- **Alerts that arrive in time.** The person has to hear about a problem while it can still be stopped.
- **A view of the work.** They need to see what the system is doing and why. That is the case for [LLM observability and monitoring](/blog/llm-observability-vs-monitoring/).
- **A stop path.** A control that interrupts the system, with a named person allowed to use it.
- **A way to undo.** If the action cannot be reversed, it belongs in the loop.
- **Sampled review.** A person checks a sample of completed actions after the fact.

Article 14 of the EU AI Act is a useful reference here. For high-risk AI systems it asks that they can be effectively overseen by natural persons while in use, and its list of oversight measures includes being able to intervene in the operation of the system or interrupt it through a "stop" button or a similar procedure. Whether a given system falls under the article is a legal question. The [text of Article 14](https://artificialintelligenceact.eu/article/14/) is worth reading with your counsel.

## Can one AI system use both?

Yes, and a system with several kinds of action should. An agent that reads a record, drafts a reply and sends it has three actions in three tiers: the read needs no control, the draft can run on the loop, and the send waits in the loop. The same split applies to [agents that act](/blog/agentic-ai-pilot-to-production/). Decide per action, write the decision down, and put it on the page that defines production, as described in [define production for an AI system](/blog/define-production-for-an-ai-system/).

## What about human out of the loop?

Out of the loop means no person is in the path of the action, before it or while it runs. It is the lightest control, and it fits a narrow set of actions: reads, and trivially reversible internal actions, with logging and a sampled review after the fact.

Anything that touches a client, moves money, changes an access right or alters a record that is hard to restore should not run out of the loop. If an action gains reach, for example by getting write access where it could only read, move it up a tier. Reads can run out of the loop with logging, reversible writes inside your operation can run on the loop, and irreversible or external actions stay in the loop.

## When can you loosen a control?

Review controls on evidence. Measure how often the approver edits or refuses, and why. When edits and refusals stay low over a period your owners agree on, propose a lighter control, such as sampled review, and decide it in writing with the same owners who set the gate. Keep the tier in mind: an action that becomes irreversible goes back behind a gate.

## What goes wrong with each model?

- **In the loop without a timeout.** The first unanswered approval stops the process, or someone starts bypassing the gate.
- **In the loop with nothing to look at.** The approver sees a button and no inputs, and approves by habit.
- **On the loop without a stop path.** The supervisor can see the problem and cannot act on it.
- **On the loop for an action that cannot be undone.** The supervision arrives after the consequence.

## What should you do next?

List the actions your AI system can take. Ask the two questions of each, and write the model beside it: in the loop, or on the loop. For every in the loop action, name an approver and a timeout rule. For every on the loop action, name who gets the alert and who can stop it. For the principle behind this, see [AI trusted by design](/blog/trusted-ai-in-operations/) and our [trust page](/trust).

Bring us the problem, not the specification. [Book a call](/contact). One call is enough to know whether we can help.

## Questions

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

In the loop, a person confirms before the action happens. On the loop, the action happens and a person supervises, with the power to stop or reverse it.

### Which one should I use for an AI system?

Choose by consequence. Use in the loop for irreversible or external actions. Use on the loop for reversible actions inside your operation, with a working stop path and a way to undo.

### Can one AI system use both?

Yes. A system can have several kinds of action. Sort each action into a tier and give each tier its own control.

### What does a person on the loop need?

Alerts that reach them in time, a view of what the system is doing, a way to stop it, and a way to undo what it did. Without those, supervision is nominal.

### Does the EU AI Act require human oversight?

For high-risk AI systems, Article 14 requires that they can be effectively overseen by natural persons while in use. Whether your system is high risk is a legal question for your counsel.

### When can I move from in the loop to on the loop?

When the record shows edits and refusals stay low, the owners agree in writing, and the action is reversible. An action that becomes irreversible goes back behind a gate.
