An organization pilots an AI system that recommends high-impact eligibility decisions. Mistakes can seriously affect individuals. Which deployment practice best supports responsible use?

1. Use only average response time to determine whether high-impact recommendations are acceptable.
2. Approve every recommendation automatically as long as the model sounds confident.
3. Require qualified human review for consequential decisions, define escalation and appeal paths, and monitor decision quality.
4. Hide the use of AI from all reviewers so they cannot question its output.