# 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.

## 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](https://learn.microsoft.com/en-us/security/benchmark/azure/mcsb-v2-artificial-intelligence-security)
- NIST SP 800-53 Rev. 5 AC-6: [source](https://csrc.nist.gov/pubs/sp/800/53/r5/upd1/final)

## Related pages

- [Question 28: AI workloads, ai activity logging](/questions/ai-activity-logging/)
- [Question 29: Guardrails and drift, preventive policies](/questions/preventive-policies/)
- [Question 30: Guardrails and drift, changes as code](/questions/changes-as-code/)

## Sources

- Microsoft cloud security benchmark AI-5: https://learn.microsoft.com/en-us/security/benchmark/azure/mcsb-v2-artificial-intelligence-security (read 16 Nov 2026)
- NIST SP 800-53 Rev. 5 AC-6: https://csrc.nist.gov/pubs/sp/800/53/r5/upd1/final (read 16 Nov 2026)

Last reviewed 16 Nov 2026. Canonical: https://cloudsecurityriskassessment.com/questions/human-approval/
