Implementation Guide

    How to Enforce US-Only Azure OpenAI Deployment Types (Global vs Data Zone vs Regional) for Customer Data

    Treat US-only as two separate tests: where inference is processed and where data is stored at rest. Permit only the deployment types that pass both under your own definition of US-only, which usually means excluding Global for customer data. Then enforce the decision in layers: Azure Policy and RBAC at provisioning, detective review of existing resources, network and endpoint restrictions, and policy enforced on each request. Retain evidence at every layer.

    A layered enforcement approach

    1. 1

      1. Restrict at provisioning

      Use Azure Policy to limit which deployment types and locations can be created, and RBAC to limit who can create or modify Azure OpenAI resources and deployments. Test each definition in a non-production subscription to confirm it blocks what you expect.

    2. 2

      2. Detect on existing resources

      Policy applied at creation does not prove that older resources comply. Run compliance scans and periodic inventory reviews of every deployment, its type and its location, and route exceptions to the named owner.

    3. 3

      3. Control the network path

      Use private networking and egress rules so workloads reach only approved resources. Microsoft describes private networking and regional availability as part of Azure OpenAI, but network placement alone does not establish where inference is processed.

    4. 4

      4. Fix client endpoint selection

      Applications choose which endpoint to call. Keep endpoints and deployment names in controlled configuration, review changes to them, and block calls to endpoints outside the approved list.

    5. 5

      5. Enforce policy on each request

      Apply rules in the request path so that a call that does not match approved policy is stopped or escalated, whatever the client was configured to do.

    6. 6

      6. Validate in production

      The NIST AI RMF describes system validation and testing in production. Run recurring acceptance tests that confirm live deployments and endpoints match the approved list.

    Evidence that shows the requirement held

    • The written definition of US-only, the approved deployment types, and the approval decision with its accountable owner
    • A current inventory of Azure OpenAI resources and deployments showing type and location, with dated review records
    • Policy assignments, compliance state over time, and the record of any approved exceptions
    • Acceptance test results showing live deployments and endpoints match the approved list
    • Network and egress configuration records showing which destinations workloads can reach
    • Request-level records of the destination and the policy decision for each interaction, retained inside your chosen residency boundary

    Start with what the deployment type actually guarantees

    Azure OpenAI Global, Data Zone and Regional deployment types differ in where inference is processed and where data is stored at rest. Those are two different properties, and a US-only requirement for customer data usually needs both to hold. A type that keeps data at rest in the US but allows processing elsewhere does not satisfy a requirement that customer data be handled only in the US. The reverse is equally true.

    The type name is only the first dimension. Residency also depends on ancillary data handling such as abuse monitoring, on whether the model and features you need are offered in the zone or region you want, and on any fallback or overflow behavior between deployments. Each of these should be settled in writing against current Microsoft documentation and your contract before the type is approved, and revisited when the documentation changes.

    The definition of US-only also belongs to your organization. Contracts, customer commitments and legal advice determine what must stay in the US. The technical decision should be traceable to that definition rather than to a deployment type label.

    Decision rules for each deployment type

    Use the table as a decision aid. Each row states the test a type must pass before it is added to the approved list for customer data.

    • Global: Approve for customer data only if both processing and data at rest are confined to the US under your definition. If processing can occur outside the US, exclude it for customer data and reserve it for non-customer workloads.
    • Data Zone (US): A candidate for US-only workloads. Approve it once the US zone scope for processing and at-rest storage matches your definition and the required model and features are available in that zone.
    • Regional (Standard): A candidate when the deployment is in a US region. Confirm model availability, capacity and ancillary data handling for that region.
    • Regional (Provisioned): Same residency test as Regional Standard, with reserved capacity as an added consideration. Confirm that any overflow or fallback path does not leave approved US scope.

    Gaps that remain after choosing Data Zone or Regional

    Selecting an approved type closes one gap and leaves others. Preventive policy only covers what it is written to cover, so confirm which properties a policy can actually evaluate, and what it does not see, such as client-side configuration. Resources that predate the policy need the detective layer. Network controls narrow where traffic can go but do not describe how the service processes it.

    Check model and feature availability for the zone or region before committing. If a required model is not offered in an approved type, teams may be tempted to create a deployment outside the approved list. Decide in advance how that exception is requested, reviewed and recorded. Also confirm that any fallback or overflow configuration cannot send requests to a deployment outside US scope.

    Legal applicability is a separate question. Regulations such as 28 CFR Part 202 on access to U.S. sensitive personal data, which as of November 2026 appears in the current eCFR and includes a Subpart D provision on restricted transactions involving a vendor, may or may not apply to your data and vendors. Have counsel determine applicability, then map the outcome to your approved-type list rather than assuming it.

    Runtime enforcement in the request path

    Provisioning controls govern what can exist. They do not govern what each request does. Enforcing policy in the execution path closes that gap by evaluating individual interactions and tool calls as they happen, so a non-conforming request is handled at runtime rather than discovered in a later review.

    Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with actions that include allow, redact, block, rate limit and escalate for human approval. In a US-only design, this layer applies your rules to each request and records the outcome. It sits alongside Azure Policy, RBAC and network controls rather than replacing them.

    Trussed is deployed self-managed in the customer cloud, with configurable retention and data residency. The evidence generated by enforcement therefore stays within the boundary you set, and it is audit-ready, with mappings to frameworks including NIST. This layer enforces and records policy decisions. It does not change how Azure processes a request once an approved call is made, so the Azure-side controls above remain necessary.

    Layers of US-only enforcement

    Decide

    Approve deployment types by processing location and data-at-rest location.

    Prevent

    Restrict deployment types, locations and who can create them.

    Contain

    Limit which endpoints clients can reach.

    Enforce and prove

    Apply policy on each request and retain evidence.

    Enforce the decision on every request

    See how policy is enforced in the execution path and how audit-ready evidence is produced in a self-managed deployment.

    Request a Demo