# Roll a guardrail out in stages, keep a dated exception path, and read the denied calls every day

Facts read 10 Nov 2026. Page reviewed 16 Nov 2026.

AWS, Azure and Google Cloud each document a staged path: test OU first for SCPs and RCPs, DoNotEnforce or Enroll for Azure Policy, dry-run for Google organization policy. This guide sets out the stages, how to handle exceptions and break-glass access, and how to monitor denied calls in CloudTrail and its equivalents.

## Stage the rollout by scope, smallest first

AWS states it strongly recommends that you do not attach an SCP to the root of your organization without thorough testing, and suggests an OU that you move accounts into one at a time, or in small numbers. For RCPs the advice is to start with individual test accounts, move up to OUs low in the hierarchy, then work up the structure.

The same ordering works elsewhere. Azure assignments and Google policies attach to a management group, folder or project, and a lower scope takes the change first. Because AWS inheritance means a deny near the root applies to everything below it, the root is the last stop and not the first. A pilot subscription or project is the cost of going slowly, and it is small.

## Learn what is used before you restrict it

Before attaching an SCP, look at what the accounts do. AWS names two sources: IAM service last accessed data, which shows which services an account used, and CloudTrail logging at the API level. A deny list for three services is safe only if the logs show none of them in use. Last accessed data shows what was used, not what will be needed next quarter, so ask the account owners as well.

For an allow list the work is larger, because every service in use must be named. AWS notes that removing FullAWSAccess without a replacement fails all actions from member accounts. Build the allow list from observed use, add a margin for global services, and test it in the OU before it goes near the root. Keep the allow list in version control beside the SCP.

## Plan exceptions and break-glass before the first deny

Most library SCPs exempt one named role with a condition. That role is the exception path, so decide who can assume it, how it is logged and how it is reviewed, before you attach the policy. A guardrail without an exception path gets removed in the first incident. Name the role in the policy exactly as it appears in the account, since a typo or an unreplaced placeholder leaves the exception empty.

Azure has a built-in version. An exemption with a Waiver category and an expiresOn date lets one resource or hierarchy stay out of the policy for a stated time, and the object stays on record after it lapses. Google organization policy can scope a rule by tag, for example enforcing a constraint only on resources tagged environment=development. Set an expiry on each exception and review them together.

## Use the report-only phase of each cloud

On Azure, assign with DoNotEnforce, read the compliance results, then switch to the default mode. When the rollout must go subscope by subscope, the Enroll mode enforces only on resources that are enrolled. On Google Cloud, set the organization policy in dry-run mode so violations are audit logged and nothing is denied, and use the dry run mode of a VPC Service Controls perimeter the same way.

A report-only switch for SCPs: none found on 2026-11-10 in the pages read, so the staged OU and the simulator take its place. Control Tower detective controls can report on a related condition in parallel. Control Tower detective controls report a status of clear or in violation during the same period, so you can watch the condition an SCP will block before it is blocked.

## Watch denied calls every day during the rollout

On AWS, AWS documents CloudTrail as the place to review Access Denied errors after an RCP is attached. On Azure, a deny returns a 403 and audit effects write to the activity log. On Google Cloud, dry-run violations land in audit logs, and Google documents Cloud Audit Logs for monitoring access patterns across VPC Service Controls perimeters. Name who reads the logs each morning during the rollout.

Set a review owner and a rollback rule before the first attachment, for example detach if production errors rise above an agreed count. Keep the previous policy JSON, since disabling the SCP policy type in an AWS root detaches every SCP and the attachments are not recoverable. [Testing a guardrail first](/guides/testing-a-guardrail-before-you-apply-it/) lists the checks that come before this stage. Write the rollback rule down before the attachment.

## A staged plan, cloud by cloud

The table puts the documented stages in one place. The names change by cloud, and the order does not: report first, smallest scope second, wider scope third. Use it as the skeleton of the change request. Fill in the owner and the date for each cell before the change is approved, and keep the table with the change record afterward.

| Stage | AWS | Azure | Google Cloud |
| --- | --- | --- | --- |
| 1. Report | Service last accessed data, CloudTrail | DoNotEnforce | Dry-run mode |
| 2. Smallest scope | Test OU or test account | Pilot subscription or Enroll | Project or folder |
| 3. Exceptions | Privileged role in the policy condition | Dated Waiver exemption | Tag condition on a rule |
| 4. Wider scope | Parent OU, then root | Management group | Parent folder, then organization |

## Related pages

- [AWS guardrails: Bedrock Guardrails vs account guardrails](/aws-guardrails/)
- [AWS Control Tower guardrails: now called controls](/aws-control-tower-guardrails/)
- [Multi cloud governance: three hierarchies, three mechanisms](/multi-cloud-governance/)
- [All of them](/guides/)

## Sources

- AWS Organizations: Service control policies: https://docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_scps.html (read 10 Nov 2026)
- AWS Organizations: Resource control policies: https://docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_rcps.html (read 10 Nov 2026)
- Azure Policy: Assignment structure: https://learn.microsoft.com/en-us/azure/governance/policy/concepts/assignment-structure (read 10 Nov 2026)
- Azure Policy: Exemption structure: https://learn.microsoft.com/en-us/azure/governance/policy/concepts/exemption-structure (read 10 Nov 2026)
- Google Cloud: Test organization policies in dry-run mode: https://cloud.google.com/resource-manager/docs/organization-policy/dry-run-policy (read 10 Nov 2026)
- Google Cloud: VPC Service Controls overview: https://cloud.google.com/vpc-service-controls/docs/overview (read 10 Nov 2026)

Last reviewed 16 Nov 2026. Canonical: https://cloudgovernancetools.com/guides/rolling-out-guardrails-without-breaking-production/
