AI governance / ZARIS resources

Human-in-the-loop AI explained

Design human review around decisions, authority and the information a reviewer needs to act.

Human-in-the-loop AI explained — editorial illustration
AI governanceZARIS Insights
The practical takeawayA review button is useful only when the person has enough context, appropriate authority and a real option to reject or change the proposed action.

Choose the decision that needs a person.

Human-in-the-loop describes a process in which people remain involved in relevant AI-supported decisions or actions. In practice, the useful question is where review belongs: before a message is sent, before a record changes or when information is uncertain.

Do not add approval steps indiscriminately. Identify the consequences of an incorrect action and the authority the system has. A low-impact draft and a sensitive account change need different controls.

Give the reviewer usable context.

Show the original request, relevant approved information, proposed action and unresolved questions. The reviewer should understand what will happen if they approve and which system will perform it.

Where an assistant suggests a category or recommendation, expose the criteria and missing information. A polished explanation is not evidence that the underlying facts are correct; the person needs access to the material required to check them.

  • Original input and record reference.
  • The proposed change or message.
  • Supporting information and uncertainty.
  • Approve, revise, reject and escalate options.

Make ownership and recovery explicit.

Name who reviews, how work is assigned and what happens if a reviewer is unavailable. Approval should be linked to the specific action version so a later change cannot inherit an earlier sign-off by accident.

The system also needs to report whether an approved action actually completed. A human approval does not remove the possibility of a failed integration, duplicate request or downstream rejection.

Evaluate the review process itself.

Check whether reviewers have enough time and information to make meaningful decisions. Track corrections, overrides, missed risks and excessive approval queues. If every item is approved without inspection, the control may be providing less assurance than expected.

NIST’s AI Risk Management Framework offers a general risk-management reference. It does not establish that a particular workflow is approved or compliant; accountability still needs to be defined for the actual deployment.

General workflow guidance. Product availability, provider requirements and implementation scope are confirmed for each project.

Further reading

Source documentation.