Decide where processing may happen and pick the deployment type that enforces it. Then check each model family's data terms, move access to Entra ID, and confirm audit logging and compliance scope before regulated data reaches Foundry.
When you need this
- Customers or regulators require processing in the US, the EU or Asia Pacific.
- Security asks whether prompts reach the model's maker or train a model.
- A public sector or regulated workload needs FedRAMP or audit coverage before launch.
Evaluation criteria
Deployment type and residency
Which deployment type keeps processing where your rules require?
Look forGlobal deployments may process in any Azure region. Data Zone deployments process only inside the US, EU or Asia Pacific zone, and Standard or regional provisioned deployments stay inside the customer-specified geography.
Data terms per model family
Is each model sold by Azure, or does a partner process the data?
Look forFor models sold by Azure, prompts and completions are not available to OpenAI or other model vendors and are not used to train foundation models without permission. Claude comes hosted on Azure or on Anthropic infrastructure, under its own terms.
Identity
Do applications and people reach Foundry through Entra ID?
Look forUse Microsoft Entra ID with token-based, keyless access, conditional access and role-based access control. Microsoft recommends Entra ID over API keys for production.
Audit and compliance
Are prompts, responses and control-plane changes logged where your auditors look?
Look forPurview Audit records prompts and responses from Foundry interactions in the unified audit log. Microsoft's FedRAMP scope page lists the Foundry portal for FedRAMP High and DoD IL2 in Azure Government.
Abuse monitoring
Do you need modified abuse monitoring?
Look forAbuse monitoring can store flagged prompts for review. Managed customers can apply for modified abuse monitoring to turn off that storage and human review.
Questions to ask vendors
- Which deployment type does each production model use today, and does it match our residency commitments?
- Which models we use are not sold by Azure, and whose data terms apply to them?
- Which applications still authenticate with API keys?
- Is Purview Audit on for every subscription that hosts Foundry?
Red flags
- Global deployments for data that must stay in the EU.
- Assuming Claude in Foundry follows the terms for models sold by Azure.
- Shared API keys across teams.
- No audit log for prompts in a regulated workload.
Products in this piece
Sources
- Official documentation
- 7
- Microsoft Learn: Deployment types for Microsoft Foundry Models
- Microsoft Learn: Data, privacy, and security for Foundry Models
- Microsoft Learn: Foundry Models from partners and community
- Microsoft Learn: Abuse monitoring
- Microsoft Learn: Authentication and authorization in Microsoft Foundry
- Microsoft Learn: Microsoft Purview for Microsoft Foundry
- Microsoft Learn: Azure services in FedRAMP and DoD audit scope