Digital Lexicon Hub

A control requiring an authorized person to approve a proposed decision or action before a system executes it.

Human approval places accountability for consequential automated actions with a named person, which reduces the risk of costly errors and supports regulatory expectations for human oversight of AI systems.

An AI agent drafts refunds for support tickets. Refunds under $50 run automatically, while larger ones enter an approval queue where a supervisor reviews the case, approves or rejects it, and the decision is logged with a timestamp.

Approval steps can become rubber-stamping when reviewers face high volume or trust the system too much, a pattern known as automation bias. Approval without enough context, time or authority to reject gives little real protection.

A human approval gate pauses execution at a defined point, presents the proposed action with supporting evidence, and resumes only after an authorized reviewer accepts, edits or rejects it. Thresholds based on risk, value or uncertainty decide which actions require review. Effective designs record who approved what and when, enforce permissions in the system rather than in prompts, and let reviewers stop or override the system.