Human review gates
Defined checkpoints where a human reviews, edits, or approves AI-generated outputs before they move forward in the workflow.
Why it matters
AI systems produce outputs that are usually right and occasionally dangerously wrong. Review gates ensure that a qualified person sees every output before it affects a decision, a report, or a communication. Without them, errors compound silently.
Where it shows up
finance
Every variance commentary draft must be reviewed by an analyst before inclusion in a reporting pack. No auto-generated number leaves the system without human sign-off.
hr
Policy guidance to managers always includes a confidence indicator. High-sensitivity responses are routed to HR for review before the manager sees them.
procurement
Vendor scoring recommendations are presented as evidence-backed suggestions, not decisions. Procurement leads review and adjust before any recommendation goes to stakeholders.
Common mistakes
- Making review optional when workload is high — the gate must be mandatory
- Reviewing only a sample of outputs instead of all outputs in high-risk workflows
- Not giving reviewers enough context to evaluate the AI output effectively
- Treating the review gate as a rubber stamp instead of a genuine quality check
Signals that a workflow needs this pattern
- The workflow produces outputs that affect financial statements, legal standing, or people decisions
- Errors in the output would be difficult or expensive to reverse
- The domain has regulatory or compliance requirements for human oversight
- Stakeholders downstream would lose trust if they discovered the output was unchecked AI