Implementation | Optional module: Azure Policy guardrails for Microsoft Foundry AI services

Chapter 5 of 6

Implement#

1. Complete the decision record#

Edit artifacts/policy/guardrail-decisions.json in the approved private path. Keep the initial effects in audit mode unless the owner has already reviewed the current policy states.

2. Run preflight#

PowerShell

$targetScope = "<approved subscription resource ID>"
$deploymentLocation = "<approved deployment location>"

.\scripts\preflight.ps1 `
  -TargetScope $targetScope `
  -DeploymentLocation $deploymentLocation

Bash

target_scope="<approved subscription resource ID>"
deployment_location="<approved deployment location>"

./scripts/preflight.sh \
  --target-scope "$target_scope" \
  --deployment-location "$deployment_location"

Preflight rejects unresolved decisions, unsupported effects, a mismatched scope, Bicep syntax errors, and a subscription mismatch. After those checks pass, it runs two subscription-scope what-if previews: one for the initiative and custom definition, and one for the assignment.

3. Deploy the initiative and staged assignment#

Deploy the initiative first:

PowerShell

az deployment sub create `
  --location $deploymentLocation `
  --name "optional-ai-services-guardrail-policy-initiative" `
  --template-file .\artifacts\policy\initiative.bicep `
  --parameters .\artifacts\environments\initiative.bicepparam `
  --only-show-errors

Bash

az deployment sub create \
  --location "$deployment_location" \
  --name "optional-ai-services-guardrail-policy-initiative" \
  --template-file ./artifacts/policy/initiative.bicep \
  --parameters ./artifacts/environments/initiative.bicepparam \
  --only-show-errors

Then deploy the assignment:

PowerShell

az deployment sub create `
  --location $deploymentLocation `
  --name "optional-ai-services-guardrail-policy-assignment" `
  --template-file .\artifacts\policy\assignment.bicep `
  --parameters .\artifacts\environments\assignment.bicepparam `
  --only-show-errors

Bash

az deployment sub create \
  --location "$deployment_location" \
  --name "optional-ai-services-guardrail-policy-assignment" \
  --template-file ./artifacts/policy/assignment.bicep \
  --parameters ./artifacts/environments/assignment.bicepparam \
  --only-show-errors

Keep the assignment in DoNotEnforce until current policy states are available and reviewed.

4. Review current compliance by reference#

PowerShell

.\scripts\check-compliance.ps1 -TargetScope $targetScope

Bash

./scripts/check-compliance.sh --target-scope "$target_scope"

The script summarizes live policy states by policy definition reference. It does not export or store live-state data in this repository.

5. Promote approved references#

Update guardrail-decisions.json for the approved promotion. Common changes are:

enforcementMode from DoNotEnforce to Default;
networkAccess, localAuthentication, or deploymentSku from Audit to Deny; and
diagnosticLogs from AuditIfNotExists to DeployIfNotExists.

Set policyStateReviewStatus to Reviewed only after the owner review. Rerun preflight and redeploy the same assignment.