Human in the Loop vs Human on the Loop: Differences, Examples and How to Choose

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

The difference is timing. In human in the loop, an AI system stops before an action and a person must approve it before anything happens. In human on the loop, the AI acts on its own while a person monitors its operation and can step in, override it or stop it when something goes wrong.

Both keep people in control of AI, but they place that control at different points. Human in the loop puts a person in the path of every decision it covers, so nothing happens without them. Human on the loop puts a person beside the process, watching outcomes and ready to intervene, so decisions flow without waiting. The diagram shows where the person sits in each model.

The human in the loop AI guide introduces these models alongside human out of the loop. This page compares human in the loop and human on the loop in depth: how each works, what each needs to be effective, how to choose between them for a given action, and how they relate to the human oversight requirements in the EU AI Act.

What is human in the loop?

Human in the loop is an oversight model in which an AI system cannot complete a decision or action until a person has reviewed and approved it. The AI does the preparation, such as drafting, scoring or recommending, but a named person has the final say every time the checkpoint is reached.

Typical examples include an AI agent that drafts a supplier payment but waits for a finance approver, a model that flags a transaction for fraud but leaves the decision to block the account with an analyst, or an agent that writes a customer reply that a team member must sign off before it is sent.

The strength of the model is that errors are caught before they have an effect. The cost is speed and capacity: every decision waits for a person, and the number of decisions the process can handle is limited by how many people are available to review them.

What is human on the loop?

Human on the loop is an oversight model in which an AI system makes and carries out decisions on its own, within agreed limits, while a person supervises its operation. The supervisor watches for anomalies, reviews exceptions and keeps the authority and the means to intervene, correct the system or stop it.

Typical examples include an agent that resets passwords or closes routine service tickets while an IT lead monitors volumes and errors, a system that adjusts stock orders within set limits while a planner reviews exceptions, or a model that routes customer enquiries while a team leader samples its decisions each day.

The strength of the model is throughput: decisions are made at machine speed and people focus on the cases that need them. The weakness is that a mistake has already happened by the time anyone sees it, so the model only suits actions whose effects can be detected and corrected.

Human in the loop vs human on the loop compared

The table compares the two models across the points that matter when designing oversight for AI agents and automated decisions.

Neither model is stronger in every respect. Human in the loop gives tighter control over individual decisions but does not scale; human on the loop scales but depends on supervisors noticing problems quickly and being able to act on them.

Where do human in command and human out of the loop fit?

Human in command and human out of the loop sit at either end of the same spectrum. Human in command is the broadest level of oversight, deciding whether and how an AI system is used at all. Human out of the loop means the system runs with no real-time human role, with people setting rules beforehand and auditing afterwards.

The European Commission's High-Level Expert Group on AI named human in the loop, human on the loop and human in command as the three approaches to human oversight in its Ethics Guidelines for Trustworthy AI (2019). France's data protection authority, the CNIL, uses the same three categories in its guidance on using AI systems in production, describing them as:

• Human in the loop: the capacity for human intervention in each decision cycle of the system.
• Human on the loop: the capacity for human intervention during the system's design cycle and in monitoring its operation.
• Human in command: the ability to supervise the overall activity of the AI system and to decide when and how to use it in a given situation.

In practice, human in command is a governance responsibility held by leadership and system owners, while human in the loop and human on the loop are operational choices made action by action. Human out of the loop is not one of the three oversight approaches; it describes the absence of real-time oversight and suits only tasks where errors are cheap and easy to reverse.

How do you choose between human in the loop and human on the loop?

Choose per action, not per system. Use human in the loop where an action is irreversible, high value, visible outside the organisation or affects someone's rights. Use human on the loop where actions are frequent, reversible and low impact, and where problems can be detected quickly. Most AI agents need both models for different actions.

Five questions settle most cases:

1. Can the action be undone? If not, such as a payment sent, an external message published or data deleted, put a person in the loop.
2. What does a single error cost? The higher the financial, legal or reputational cost, the stronger the case for approval before the action.
3. How many decisions are there? If the volume would overwhelm reviewers, approval will turn into rubber-stamping. Narrow the checkpoint or move lower-risk cases on the loop.
4. How quickly must the action happen? If a delay of minutes causes harm, such as isolating a compromised device, human on the loop with fast alerts and a stop control may protect people better than a queue.
5. Does a rule or policy require a named approver? Internal delegations of authority, maker-checker controls and some regulations require a person to sign off specific decisions.

Applied to common AI agent actions, the answers often look like this:

These are starting points, not rules. The right answer depends on your thresholds, your regulators and how well each agent has performed so far.

What each model needs to work in practice

Choosing a model is only half the decision. Each one fails in a predictable way if the human step is not designed properly.

Human in the loop needs requests that are seen and answered

• A named approver. Each request goes to a specific person or a defined group, not a shared channel that everyone can see and nobody owns.
• Enough context to decide. The request states what the AI wants to do, why, and what happens if it is approved.
• A deadline and an escalation path. If the approver does not respond in time, the request goes to a backup. It should never approve itself by default.
• A manageable volume. If approvers routinely approve almost everything unchanged, the checkpoint is too broad.
• A record. The request, the approver, the decision and the timestamps are kept as evidence.

Human on the loop needs signals that cut through

• Clear limits. The AI acts only within boundaries agreed in advance, such as value thresholds or permitted systems.
• Exception alerts, not raw feeds. Supervisors need to be told when something looks wrong, such as unusual values, error spikes or complaints, rather than watching every action.
• Alerts that reach the right person quickly. An anomaly that sits unread in a dashboard or inbox is not oversight.
• A working stop control. Supervisors must be able to pause or halt the system and know how to do it.
• Regular sampling. Reviewing a sample of normal decisions catches slow drift that never triggers an alert.

Both models also share a human risk. People who oversee a system that is usually right tend to defer to it, which the EU AI Act calls automation bias. Rotating reviewers, training them on how the system fails and tracking override rates all help keep oversight real. For more on why delivery is the weak point in many designs, see why Slack and email are not a human-in-the-loop strategy for AI agents.

Does the EU AI Act require human in the loop or human on the loop?

The EU AI Act does not mandate either model by name. Article 14 requires high-risk AI systems to be designed so that people can oversee them effectively while in use, with measures proportionate to the risks, level of autonomy and context. That can be met by human in the loop, human on the loop or a combination.

Under Article 14 of Regulation (EU) 2024/1689, the people assigned to oversight must be enabled, as appropriate and proportionate, to:

• understand the system's capacities and limitations and monitor its operation, including detecting anomalies
• remain aware of automation bias
• correctly interpret the system's output
• decide not to use the system in a particular situation, or disregard, override or reverse its output
• intervene in its operation or interrupt it through a stop button or similar procedure

Most of these describe human on the loop capabilities: monitoring, overriding and stopping. The Act goes further for one category. For remote biometric identification systems listed in point 1(a) of Annex III, no action or decision may be taken on the basis of an identification unless it has been separately verified and confirmed by at least two people with the necessary competence, training and authority, subject to limited exceptions. That is a human in the loop requirement in all but name.

Whether a particular system is high risk, and which measures are appropriate, depends on its use and should be confirmed with legal advice. Oversight design is one part of compliance, not the whole of it.

Moving actions between the two models

The choice is not permanent. Many organisations start a new AI agent with human in the loop for most actions, then move specific actions on the loop as evidence builds that the agent handles them well.

• Start narrow. Begin with approval on every consequential action while you learn how the agent behaves.
• Measure. Track how often approvers amend or reject each type of request. A long run of unchanged approvals for one action type is a sign it may be ready to move on the loop.
• Move one action type at a time, with limits and exception alerts in place before approval is removed.
• Keep a route back. If error rates, complaints or alert volumes rise, return the action to human in the loop until the cause is fixed.
• Record each change to the oversight model and the evidence behind it, so the decision can be explained later.

For a wider view of how oversight shifts as AI use matures, see the three stages of AI adoption.

Where Heed fits in both models

Heed is internal communication software built to make sure important messages are seen, acknowledged and recorded. Both oversight models depend on that: human in the loop needs approval requests that reach a named person, and human on the loop needs exception alerts that a supervisor cannot miss.

• Triggered by your systems. Desktop alerts can be triggered via API and webhooks, or from ServiceNow, Splunk, SCOM and Jira, so an agent or monitoring tool can raise a request or an exception alert automatically.
• Sent to the right people. Alerts can be targeted by department, location or role through Active Directory and SAML sync.
• Answered from the alert. For in-the-loop approvals, alerts can include action buttons so approvers can respond from the notification, and unanswered requests can escalate to a secondary approver or manager, as described on our approvals solution page.
• Acknowledged and recorded. Acknowledgement can be required before a high-priority alert is dismissed, giving a named, timestamped record of who confirmed it. Our guide to audit-ready communication explains why that evidence matters.
• Deployed where your data must stay. Heed offers true on-premises deployment as well as regional cloud options, and is ISO 27001 certified.

Heed does not decide which actions need approval, does not monitor AI systems itself and does not make an AI system compliant on its own. It provides the dependable human communication step that an oversight design relies on. Book a demo to see how approval requests and exception alerts work in practice.