human in the loop AI agents guardrails

Human-in-the-Loop & Guardrails

Place people where judgment changes risk, and give them evidence they can review quickly.

DIFFICULTY
Advanced
EST. TIME
55 min
UPDATED
2026-08-19
On this page
  1. 01Definition
  2. 02Why it matters
  3. 03Step by step
  4. 04Build it
  5. 05Failure modes
  6. 06Production notes
  7. 07Sources

Learning objectives

  • Choose approval, review, sampling, and escalation controls
  • Design review interfaces around changed fields and evidence
  • Prevent human approval from becoming an unmeasured rubber stamp

Prerequisites

A task risk classification · Named process owner · Evaluation rubric

Answer first

Canonical definition

Human-in-the-loop design assigns a person a specific decision, review, labeling, or escalation role at a defined point in an AI workflow, with enough evidence and authority to change the outcome.

Operating context

Why it matters

Adding a human does not make a system safe if the review arrives too late, lacks context, or is expected to process an impossible volume.

The right control depends on impact, reversibility, confidence, and the cost of delay.

Method

Step by step

  1. 01

    Classify decisions

    Separate low-impact recommendations, reversible actions, external communications, regulated decisions, and irreversible changes.

  2. 02

    Choose the control

    Use pre-approval, post-review, sampling, dual control, or mandatory escalation based on risk.

  3. 03

    Build the review packet

    Show proposed action, changed fields, evidence, uncertainty, policy checks, cost, and available alternatives.

  4. 04

    Measure reviewer behavior

    Track approval rate, time, disagreement, corrections, missed errors, and signs of automation bias.

Hands-on lab

Design a high-signal approval step

Create a review screen for an agent-proposed external email or account change.

Deliverable

A decision rubric, evidence layout, approve/edit/reject paths, timeout behavior, and audit record.

What breaks

Common failure modes

  • F1Sending every action to review until the queue becomes a bottleneck
  • F2Showing only polished output without the evidence and changed state
  • F3Treating an approval click as proof the reviewer understood the decision

Beyond the demo

Production notes

  • Calibrate review load with sampling and risk tiers, and rotate reviewers where fatigue matters.
  • Preserve the original proposal, reviewer edits, reason code, identity, and final executed action.

Further reading

Sources

  1. [1]NIST AI Risk Management Framework

Related Tenten resources

From capstone to production

A working demo is the beginning of the engineering problem.

Tenten helps teams scope one agentic workflow, integrate real data and permissions, define evals and handoffs, and own the path through launch and on-call.