Do destructive or high-impact agent actions (delete, grant access, pay, send externally) need a human approval?
Facts read 16 Nov 2026. Page reviewed 16 Nov 2026.
Question 27 of the assessment asks: Do destructive or high-impact agent actions (delete, grant access, pay, send externally) need a human approval? Approval is the last control when an agent's reasoning or inputs are wrong. It comes from Microsoft cloud security benchmark AI-5 and carries weight 2 in the ai workloads domain.
Start the assessment ยท All 40 questions
Why it matters
Approval is the last control when an agent's reasoning or inputs are wrong.
Answers and what each is worth
| Answer | Value |
|---|---|
| Yes, enforced in the tool | 1 |
| Documented, not enforced | 0.5 |
| No or not known | 0 |
How to check it in each cloud
- AWS
- Check agent action groups for user confirmation settings and that Step Functions or Lambda workflows include a manual approval for high-impact calls.
- Azure
- Check that agent tools call approval workflows (Logic Apps approvals or Power Automate) before high-impact actions.
- Google Cloud
- Check that agent tools call approval workflows or Cloud Workflows callbacks before high-impact actions.
The fix
Require an approval step in the tool or workflow for each destructive or high-impact action. Effort: medium.
Where it comes from
- Microsoft cloud security benchmark AI-5: Ensure human-in-the-loop
- NIST SP 800-53 Rev. 5 AC-6: source
Related pages
Sources
- Microsoft cloud security benchmark AI-5, read 16 Nov 2026
- NIST SP 800-53 Rev. 5 AC-6, read 16 Nov 2026
Last reviewed 16 Nov 2026.