liveBook · Manning

Human-in-the-loop

As explained in AI Governance

An approach in which a person participates in an AI system’s operation, such as by reviewing or authorizing its outputs or actions.

A clinician who accepts a fabricated AI-generated note without checking it is technically in the workflow, but is not exercising meaningful oversight. Human-in-the-loop (HITL) places a person in an AI workflow to review an output or approve an action before it proceeds. The reviewer needs time, context and authority to challenge the system and change the outcome. section 5.4.2 section 4.7.1

HITL differs from human-on-the-loop oversight, where the system can act while a person monitors it and can intervene. Some operations are too consequential to entrust to AI execution at all; define those exclusions explicitly rather than relying on approval prompts alone. section 4.7.1

Design review around the decisions and failure modes that matter:

  • Route sensitive outputs and actions for review, such as patient communications, claim denials, payments or changes to user data. Document each sensitive tool’s permissions and approval checkpoints. section 5.4.3 section 2.6.2
  • Make review actionable. Show the interpretation and critical parameters; for high-impact actions, require verification of each critical parameter before execution. section 4.7.1
  • Set thresholds to avoid approval fatigue. Reviewing every routine agent action can make the workflow unusable and train people to approve without reading. section 4.8.6
  • Route cases to a person when objective checks flag uncertainty, such as missing required fields or contradictory retrieved documents. section 6.4.1

A claim decision

An insurance workflow can pause before a denial letter, fraud flag or account closure and route the case to a human adjuster when the claim exceeds a threshold or needs extra care. Reviewers record whether they agree, partially agree or disagree with the AI suggestion and briefly explain why. Recurring reviews of overrides can expose retrieval gaps, prompt weaknesses or document-parsing errors. section 6.2.4 section 6.4.1

Human review is one layer, not a substitute for technical containment. Limit tools, action rates and scope; log decisions and provide a way to stop an agent that goes off course. Monitor outbound requests separately: a seemingly read-only URL fetch can transmit sensitive data. section 2.6.4 section 4.7.1 section 4.8.6

Track override patterns and examine incidents after deployment. A review gate should be built into the workflow so it cannot be skipped, and reviewers should be encouraged to question outputs rather than defer automatically. section 6.8 section 6.0

How AI Governance covers it

How deep it goes
Details review triggers, approval thresholds, reviewer authority and supporting controls, while recognizing that oversight of rapid black-box agents remains unsolved.
Examples
Uses a fabricated clinical note and an insurance claim workflow to illustrate rubber-stamping and targeted review.
What you can do after
Operate/configure review workflows by selecting approval gates, giving reviewers authority and using override patterns to identify system weaknesses.
How it connects
Connects human oversight to automation bias, audit logging, access control and agentic AI.

Also explained in

Compare how 28 books explain human-in-the-loop →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.