Skip to main content

EPC Group designs governed Azure OpenAI deployments for enterprises in regulated industries that need prompts, data and spend kept inside their own Azure subscription. A governed deployment is six decisions made in order: where the resource lives and which deployment type processes prompts; how the network is closed to the public internet; how content filters and abuse monitoring are configured; how retrieval keeps answers inside company data the user is allowed to read; whether capacity is standard or provisioned; and who holds which roles, what is logged, and what the budget is.

Key Facts

  • Prompts, completions and training data are not available to OpenAI, are not used to train foundation models without the customer’s instruction, and the models do not interact with OpenAI-operated services (Microsoft Learn, Data, privacy, and security for Models sold by Azure in Microsoft Foundry, accessed 2026-11-22).
  • Global deployments may process prompts in any geography where the model is deployed; Data Zone deployments stay within the US, EU or Asia Pacific zone; data at rest stays in the customer-designated geography (same source).
  • Public network access can be disabled so private endpoints are the exclusive path to the resource (Microsoft Learn, Configure Azure OpenAI networking, accessed 2026-11-22).
  • Default content filters block medium and high severity in four harm categories; prompt filtering can be set to annotate-only or off by any customer, while turning filtering off on completions requires Microsoft approval (Microsoft Learn, Configure content filters, accessed 2026-11-22).
  • Provisioned throughput reserves capacity in PTUs, with quota granted per subscription, region and deployment type; standard deployments bill per token (Microsoft Learn, What is provisioned throughput?, accessed 2026-11-22).

Last updated by Errin O'Connor, Founder & Chief AI Architect, EPC Group

Reviewed by Errin O'Connor, Founder & Chief AI Architect

What a governed deployment is, and what Microsoft already covers

An enterprise asking what a governed Azure OpenAI deployment looks like is really asking which decisions have to be made before the first production prompt is sent, and in what order. EPC Group, a Houston-based Microsoft consulting firm, builds these deployments inside the client's own Azure subscription, and the six decisions below are the order the firm's architects follow. Microsoft now documents the service as Azure OpenAI in Microsoft Foundry Models, and the controls on this page apply to the models Microsoft designates as sold by Azure (Microsoft Learn, Data, privacy, and security for Models sold by Azure in Microsoft Foundry, accessed 2026-11-22).

Microsoft states that prompts, completions, embeddings and training data are not available to other customers, are not available to OpenAI or other model providers, and are not used to train foundation models without the customer's permission or instruction; the models are hosted in Microsoft's Azure environment and do not interact with services operated by OpenAI (same source). Governance is everything the customer must still add: network, filters, data access, capacity and operations.

Decision 1: deployment type and region

The first choice is where prompts are processed. Microsoft offers Standard deployments billed per call within a region, Global deployments that route traffic across Azure regions for higher throughput, Data Zone deployments bounded to a zone such as the United States or the European Union, and Provisioned deployments with reserved capacity (Microsoft Learn, Deployment types for Azure AI Foundry Models, accessed 2026-11-22). For Global deployment types, prompts and responses may be processed in any geography where the model is deployed; for Data Zone types, processing stays inside the specified zone; in both cases data at rest stays in the customer-designated geography (Microsoft Learn, Data, privacy, and security for Models sold by Azure in Microsoft Foundry, accessed 2026-11-22). A regulated client whose policy names a processing location therefore picks Standard or Data Zone and uses Azure Policy to block the Global SKU (Deployment types page, same access date). Model support differs by region and deployment type, so EPC Group checks Microsoft's model availability by deployment type and region for the exact version the application needs before fixing the region (Microsoft Learn, Understanding deployment types in Microsoft Foundry Models, accessed 2026-11-22).

Decision 2: closing the network

Securely deploying Azure OpenAI starts with removing the public endpoint. On the resource's Networking tab, access is set to Disabled so that no network can reach the resource except through private endpoint connections; the private endpoint targets the account sub-resource and integrates with a private DNS zone named privatelink-openai-azure-com (Microsoft Learn, Configure Azure OpenAI networking, accessed 2026-11-22). For an application that grounds answers on company documents, Microsoft's reference design places Azure OpenAI, Azure AI Search and a storage account behind three private endpoints in one virtual network; selected-network IP rules are not supported because the services' addresses are dynamic, and the services authenticate to each other with managed identities and role assignments rather than keys (Microsoft Learn, Network and access configuration for Azure OpenAI On Your Data, accessed 2026-11-22; that reference design was written for On Your Data, which Microsoft retired on October 14, 2026).

Decision 3: content filters and abuse monitoring

Content filtering runs alongside the model on both the prompt and the completion. Microsoft classifies four harm categories, violence, hate, sexual and self-harm, at four severity levels, and the default configuration blocks medium and high severity for both directions. All customers can tighten or loosen thresholds per category and separately for prompts and completions; any customer can set prompt filtering to no filtering or annotate-only, while doing so for completions requires approval through Microsoft's Limited Access review, and Prompt Shields for direct attacks and protected-material detection are on by default (Microsoft Learn, Configure content filters, accessed 2026-11-22). Abuse monitoring is separate: the service applies automated review to detect recurring patterns that may violate the code of conduct, and when indicators appear a sample of prompts and completions may be stored, within the customer's geography, for review by authorized Microsoft employees (Microsoft Learn, Data, privacy, and security for Models sold by Azure in Microsoft Foundry, accessed 2026-11-22).

Decision 4: grounding on governed data

Retrieval-augmented generation is how the deployment stays inside company knowledge: each prompt is augmented with passages retrieved from a governed index before the model answers. Azure OpenAI On Your Data, the managed feature that did this, retired on October 14, 2026; Microsoft points to Foundry IQ, and the migration is not automatic (Microsoft Learn, Network and access configuration for Azure OpenAI On Your Data, accessed 2026-11-22). In Azure AI Search, security filters match a groups field in the index to the signed-in user's groups so search returns only documents the user may read, but the principal is only a string with no authentication behind it; for document-level authorization without that application-managed pattern, Azure AI Search also offers built-in access control list support (Microsoft Learn, Security Filter Pattern, accessed 2026-11-22). EPC Group's rule for clients already running Microsoft 365 Copilot is that the custom index must never be more permissive than the SharePoint sites it was built from.

Decision 5: standard or provisioned capacity

Standard deployments suit low-to-medium volume with bursts. Provisioned throughput is for well-defined, predictable throughput and latency requirements, typically once an application is in production; capacity is sized in provisioned throughput units, quota is granted per subscription, region and deployment type, and when a provisioned deployment is fully used it returns non-200 responses such as HTTP 429, which optional spillover can route to a standard deployment in the same resource (Microsoft Learn, Provisioned throughput for Foundry Models, accessed 2026-11-22). EPC Group's practice is to launch on Standard, measure tokens per minute for a full business cycle, and move to provisioned only when the measured pattern justifies a reservation.

Decision 6: roles, logging and cost controls

Access to deploy and to call models runs through Microsoft Entra roles and managed identities; Azure Policy can block deployment types that are not approved. Any data a stateful feature stores, such as Responses API message history or uploaded files, is held in the customer's Foundry resource, encrypted at rest with the option of customer-managed keys, and deletable by the customer at any time (Microsoft Learn, Data, privacy, and security for Models sold by Azure in Microsoft Foundry, accessed 2026-11-22). EPC Group adds central diagnostic logging, an API gateway with per-application quotas, and budget alerts per deployment.

Azure OpenAI or ChatGPT Enterprise for a regulated company

The question is less about model quality than about contract and boundary. Azure OpenAI runs inside the customer's Azure subscription under the Microsoft Products and Services Data Protection Addendum, inherits Entra identity, private networking and Azure Policy, and lets the customer choose the processing geography through the deployment type (Microsoft Learn, Data, privacy, and security for Models sold by Azure in Microsoft Foundry, accessed 2026-11-22). EPC Group does not characterize OpenAI's own enterprise terms; a regulated buyer should read them and compare on four criteria: where prompts are processed, whose data processing agreement governs, whether the service can be placed behind the organization's private network, and how it integrates with existing Purview labels and audit logs.

Documented work

Anonymized engagement: EPC Group engagement data in financial services and manufacturing covers mid-size and large organizations that deployed Azure OpenAI with private endpoints, document-level access control and central logging, delivered by the firm's Azure consulting practice. Client names are not published here. Engagements are described by industry and organization size only, and no outcome figures are quoted that cannot be traced to a signed deliverable. EPC Group is a Microsoft Solutions Partner with all six Solutions Partner Designations as of October 2026.

Primary sources

Frequently Asked Questions

How is Azure OpenAI isolated from the public internet (private endpoints, VNet)?

Set public network access to Disabled on the resource's Networking tab so private endpoint connections are the only way in; the private endpoint uses resource type Microsoft.CognitiveServices/accounts and the privatelink-openai-azure-com DNS zone (Microsoft Learn, Securing Azure OpenAI inside a virtual network with private endpoints, accessed 2026-11-22). Search and storage get their own private endpoints; IP rules are unsupported because service addresses are dynamic (Microsoft Learn, Network and access configuration for Azure OpenAI On Your Data, accessed 2026-11-22).

What are content filters and abuse monitoring, and can they be adjusted?

Content filters classify prompts and completions for violence, hate, sexual and self-harm at four severity levels, and the default filters medium and high. Any customer can change thresholds separately for prompts and completions and can set prompt filtering to annotate-only or off; turning filtering off on completions needs Microsoft approval (Microsoft Learn, Configure content filters, accessed 2026-11-22). Abuse monitoring may select flagged prompts and completions for automated and human review (Microsoft Learn, Data, privacy, and security, accessed 2026-11-22).

How does retrieval-augmented generation keep answers inside company data?

Passages are retrieved from a governed index and added to the prompt before the model answers. Azure OpenAI On Your Data retired on October 14, 2026, and migration to Foundry IQ is not automatic (Microsoft Learn, accessed 2026-11-22). In Azure AI Search, a security filter matches a groups field to the signed-in user's groups, but the principal is just a string: the application must authenticate users and apply the filter to every query (Microsoft Learn, Security Filter Pattern, accessed 2026-11-22).

Global Standard, Data Zone or Standard: where are prompts processed?

Global types may process prompts in any Azure region; Data Zone types process them only inside the Microsoft-defined US, EU or Asia Pacific zone; Standard and Regional Provisioned keep processing inside the customer-specified Azure geography. Data at rest stays in the designated geography for every type, and Azure Policy can block a type such as GlobalStandard (Microsoft Learn, Understanding deployment types in Microsoft Foundry Models, accessed 2026-11-22). Not every model supports every type, so check availability by region first.

Provisioned throughput (PTU) or pay-as-you-go?

Start on pay-per-token Standard deployments for development, testing and variable traffic, and move to provisioned throughput when traffic is predictable, latency-sensitive and large enough that per-token billing becomes expensive. Capacity is bought in provisioned throughput units, quota is granted per subscription, region and deployment type, and quota does not guarantee capacity (Microsoft Learn, Provisioned throughput for Foundry Models, accessed 2026-11-22). EPC Group measures a full business cycle on Standard before recommending a reservation.

What does governance of an Azure OpenAI service require (roles, logging, cost controls)?

Governance covers roles, logging and spend. Services authenticate to each other with managed identities and role assignments, and Azure Policy can block unapproved deployment types (Microsoft Learn, accessed 2026-11-22). Data stored by stateful features such as the Responses API stays in the customer's Foundry resource, encrypted at rest with the option of customer-managed keys, and deletable by the customer at any time (Microsoft Learn, Data, privacy, and security, accessed 2026-11-22). EPC Group adds logging, an API gateway and budget alerts.

Next step

If your organization is planning an Azure OpenAI deployment under regulatory or audit requirements, request a scoping call and a senior EPC Group architect will walk through the six decisions above against your subscription. The work is delivered by the AI governance consulting practice; each engagement is a fixed fee quoted after a scoping call.

Email contact@epcgroup.net or call 888-381-9725.

a Microsoft Gold Partner (2000 until Microsoft retired the program in 2022) — today holding all six Microsoft Solutions Partner Designations.

AI assistant — not a human