Do this
- List actions you can’t take back: sending, paying, deleting and publishing.
- Set AI tools to ask a person first. Instead of letting an assistant send replies itself, have it save drafts for approval.
- Review approvals on purpose, not out of habit.
Sources
- IMDA Singapore, Model AI Governance Framework for Agentic AI, Version 1.5 (2026), section 2.2.2, human approval before irreversible actions such as deleting data, sending communications and making payments (p. 29)
- NIST AI Risk Management Framework 1.0 (NIST AI 100-1), MANAGE 4.1, override mechanisms
