# Block what cannot be undone, report what can, and use a report-only phase to learn which is which

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

A preventive control disallows the action; a detective control finds the noncompliant resource afterward; a proactive control checks a resource before deployment. AWS Control Tower, Azure Policy and Google Cloud each document all three. This guide sets out the differences, the cost of each, and a rule for choosing.

## Preventive controls stop the request, detective controls report afterward

AWS Control Tower defines a preventive control as one that keeps accounts compliant because it disallows actions that lead to policy violations. Its status is enforced or not enabled, and it is implemented with service control policies or resource control policies. A detective control finds noncompliance of resources and alerts through a dashboard; its status is clear, in violation or not enabled.

The practical difference is timing. A preventive control fails a request and the engineer sees an error. A detective control lets the change land and surfaces it later, and the resource exists in the meantime. The [preventive control](/glossary/preventive-control/) and [detective control](/glossary/detective-control/) glossary entries have the exact terms. Each has a cost that the other avoids, and a mature program runs both on the same rule.

## A proactive control checks the template before anything is created

AWS Control Tower documents a third kind, the proactive control, which checks resources before they are deployed. It is implemented as CloudFormation hooks that run on preCreate and preUpdate. AWS notes proactive controls may not affect requests made directly through the console, the API, the SDKs or other infrastructure-as-code tools. AWS sorts proactive controls into four categories, and the reference guide groups them by AWS service.

That scope limit is why a proactive control is not a substitute for a preventive one. It catches mistakes early in a pipeline that uses CloudFormation, and it leaves every other path open. Pair it with an SCP or RCP when the rule must hold on every path, and with a detective control when you need to see what already exists.

## The same split exists in Azure and Google Cloud

Azure Policy expresses it as an effect. Deny returns a 403 before the request reaches the resource provider. Audit creates a warning in the activity log and marks the resource non-compliant without stopping the request. The enforcement mode DoNotEnforce goes further, and evaluates the assignment without triggering the effect. Microsoft also documents that auditIfNotExists and deployIfNotExists are often interchangeable, which gives a path from detecting to correcting.

Google Cloud uses dry-run mode: an organization policy in dry-run is enforced like any other, but violations are audit logged and the violating actions are not denied. VPC Service Controls offers the same for perimeters. In all three clouds, the detective setting is the first step toward the preventive one. Dry-run is available only for the constraint types Google lists, so check support first.

## Choose by what an undo costs

Block when the action cannot be undone or the damage lands before anyone can react: leaving the organization, scheduling KMS key deletion, disabling logging, making a snapshot public, running a workload in a Region you have not approved. The library marks these in the [SCP mechanism](/mechanisms/scp/) and [declarative policy](/mechanisms/declarative-policy/) pages. In each case the harm is done before a report could arrive.

Report when the state is fixable and the rule is still being learned: tag conventions, naming, a setting that varies by team. A detective control costs a finding to triage, and a preventive one costs a failed deployment and a support ticket. A rule that has not run in report mode for one cycle is a risk to ship as a block.

## Run every new rule detective first, then promote it

The promotion path is documented by all three providers: AWS recommends testing an SCP in an OU before the root, Azure documents DoNotEnforce for validating logic before enforcement, and Google documents dry-run to monitor how a policy change affects workflows before it is enforced. Each of the three is a staging step, and none is a permanent state. Skipping the step is the usual cause of a rollback.

Keep the detective control after promotion. A deny stops the paths it covers, and a detective control finds what slipped past through the management account, a service-linked role or an exempted scope. The [rollout guide](/guides/rolling-out-guardrails-without-breaking-production/) turns this into stages. Reviewing detective findings monthly also catches drift in the exception list, since each exempted account or role is a place the block does not reach.

## The terms line up across the three clouds

The table maps the documented names. Providers use different words for the same three roles, and a rule is easiest to move between clouds when you map it by role first. Use the row for your cloud when you plan the promotion path. The Azure column has one gap: a pre-deployment check was not found in the pages read on 2026-11-10, so an Azure pipeline check has to be added by other means.

| Cloud | Preventive | Detective or report-only | Before deployment |
| --- | --- | --- | --- |
| AWS | SCP, RCP, declarative policy | Control Tower detective controls on AWS Config rules | Proactive controls as CloudFormation hooks |
| Azure | Policy deny effect, deny assignment, lock | Policy audit effect, DoNotEnforce mode | Not found in the pages read |
| Google Cloud | Organization policy, IAM deny, perimeter | Dry-run mode, perimeter dry run | Policy Simulator |

## Related pages

- [IAM Access Analyzer as a guardrail tool: before and after a change](/guides/iam-access-analyzer-as-a-guardrail-tool/)
- [Testing a guardrail before you apply it: simulators and dry-run modes](/guides/testing-a-guardrail-before-you-apply-it/)
- [Rolling out guardrails without breaking production](/guides/rolling-out-guardrails-without-breaking-production/)
- [All of them](/guides/)

## Sources

- AWS Control Tower: Preventive controls: https://docs.aws.amazon.com/controltower/latest/controlreference/preventive-controls.html (read 10 Nov 2026)
- AWS Control Tower: Detective controls: https://docs.aws.amazon.com/controltower/latest/controlreference/detective-controls.html (read 10 Nov 2026)
- AWS Control Tower: Proactive controls: https://docs.aws.amazon.com/controltower/latest/controlreference/proactive-controls.html (read 10 Nov 2026)
- Azure Policy: Definition effects basics: https://learn.microsoft.com/en-us/azure/governance/policy/concepts/effect-basics (read 10 Nov 2026)
- Azure Policy: Assignment structure: https://learn.microsoft.com/en-us/azure/governance/policy/concepts/assignment-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)

Last reviewed 16 Nov 2026. Canonical: https://cloudgovernancetools.com/guides/preventive-vs-detective-controls/
