85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What is human in the loop (HITL)?

Human in the loop is a design pattern in which an AI system pauses at defined points, such as high-risk actions or low-confidence decisions, so a person can approve, edit or reject the step before it takes effect.

An agent that issues refunds, edits records or sends messages will sometimes be wrong. Human in the loop puts a person at the points where a mistake would be expensive or hard to undo, and lets the agent run freely everywhere else.

The hard part is placement. Too few approvals and a bad action slips through. Too many and people start approving without reading, which is less safe than it looks.

Where approvals belong

  • High-risk or irreversible actions

    Payments above a limit, deletions, external messages and changes to permissions.

  • Repeated failure

    When the agent fails a check more than a set number of times, it stops and hands over.

  • Low confidence or out of policy

    Cases the agent cannot classify, or that fall outside the rules it was given.

  • Sensitive people and topics

    Complaints, vulnerable customers, legal threats and anything the policy names.

What the labs say

OpenAI's practical guide to building agents plans for human intervention when failure thresholds are exceeded or actions are high-risk. Google lists human in the loop as a named pattern in its agent design catalogue, and its ADK guidance treats approval gates as pauses: the agent's state is saved and the run resumes when the person responds, even days later.

Anthropic warns about approval fatigue: frequent prompts lead people to rubber-stamp them. Its research on agent autonomy found that experienced users approve more automatically but interrupt more often, as oversight shifts from approving each step to monitoring and stepping in.

How Motiontic's engineers apply it

Motiontic's engineers rate each action an agent can take by impact and reversibility, and put approvals only where the rating calls for one. The approval shows the person what the agent wants to do, why, and what it looked at, so the decision takes seconds rather than a fresh investigation.

Approval rates and overrides are tracked. When a type of step is approved without change for long enough, the team can choose to move it to monitoring instead.

Human in the loop compared with human on the loop

Human in the loop compared with human on the loop
Human in the loopHuman on the loop
Person's roleApproves selected steps before they runMonitors the running system and intervenes
TimingBefore the actionDuring or after the action
FitsHigh-risk, irreversible or regulated stepsHigh-volume, low-risk, reversible steps
Main riskApproval fatigue and slow throughputA bad action runs before anyone sees it

Common questions

What does human in the loop mean in AI?

It means an AI system is designed to stop at chosen points and wait for a person to approve, change or reject what it is about to do, usually for high-risk actions or uncertain decisions.

What is approval fatigue?

Approval fatigue is when people see so many permission prompts that they approve them without reading. Anthropic cites it as a reason to approve by risk tier and use sandboxes, rather than prompting on every step.

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

Human in the loop requires a person's approval before a step runs. Human on the loop lets the system act while a person monitors and steps in when something looks wrong.

Does human in the loop slow agents down?

Only at the gates. Agents run freely on low-risk steps, and long-running designs save their state at an approval gate and resume when the person responds.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.