Skip to main content

​
Azure AI Foundry — Inference on your account

With this integration, SecureAI’s large language model (LLM) calls run on your own Azure AI Foundry resource instead of on SecureAI-managed providers. Usage is billed by Microsoft to your Azure subscription, so it draws down your Azure commitment (MACC) and credits. SecureAI stays in the path of every request: DLP, model policies, data residency, and signed receipts apply exactly as they do with any other provider.
Not to be confused with the other Azure card. This page connects Foundry as an inference provider (Admin → Integrations → AI providers). The Azure AI Foundry — Discovery card (Cloud category) inventories agents, deployments, identities, and costs in read-only mode and never sends a prompt; it is described in Azure AI Foundry. You can use both on the same subscription, but they are independent and ask for different permissions.

​
What runs on Foundry and what does not

Supported model families: OpenAI (GPT‑5 and o‑series), gpt‑oss, DeepSeek, Llama, Grok, and Mistral — any of them that you have deployed on the resource. Claude on Foundry is not supported yet (it uses Anthropic’s API, not the OpenAI one).

​
How many cloud providers you can use

With more than one provider enabled, each one serves different model families: a family can be routed to only one provider, and only one provider can use Remap everything. SecureAI refuses a configuration that overlaps another one and tells you which family collides.

​
Prerequisites

  • An Azure AI Foundry (or Azure OpenAI) resource with at least one model deployment (for example gpt-5.1 or gpt-oss-120b).
  • A service principal (app registration in Microsoft Entra ID) with a client secret and, on that resource:
    • Cognitive Services OpenAI User — to call the models (required).
    • Reader — so SecureAI can list your deployments and map models automatically (recommended).
  • Admin permission on the Integrations section of SecureAI.

​
Step 1. Prepare the service principal in Azure

From Azure Cloud Shell or with the az CLI. Replace MY_RESOURCE and MY_RG:
Step 2 prints appId (client ID), password (client secret), and tenant (directory ID). You will paste them in SecureAI.
Treat the client secret like a password. SecureAI stores it encrypted and never shows it again. Client secrets expire (one year by default with az): put a reminder to rotate it, or the integration stops working that day.
Roles assigned in Azure can take up to 5 minutes to take effect. If the first connection test returns Missing role, wait and test again.

​
Step 2. Connect the resource in SecureAI

  1. Go to Admin → Integrations → AI providers and open Azure AI Foundry — Inference on your account.
  2. Under Connection, fill in:
  3. Click Save. SecureAI lists your deployments (name, model, version, and deployment type) and maps each catalog model to the deployment that serves it — for example, openai/gpt-oss-120b → your gpt-oss-120b deployment. Nothing is routed yet: the integration is saved turned off so you can test it first.
If the deployment list shows a warning, the principal does not have Reader on the resource, or the subscription / resource group is missing. Inference still works; you just have to type the deployment names yourself.

​
Step 3. Choose what runs on Foundry

  • Only these model families — the families you tick run on your Foundry; everything else keeps running where it did. Default: OpenAI.
  • Remap everything to Foundry — all LLM calls run on Foundry; models with no deployment are served by a default deployment.
Default deployments (pick from your deployments): Under Model overrides you can send a catalog model to a specific deployment (for example openai/gpt-5.2 → prod-chat). Overrides win over the automatic mapping. Fallback and user quotas work as on Vertex AI: with fallback off (recommended) a request that Foundry cannot serve is refused, never sent to another provider.

​
Step 4. Test, review, and turn it on

  1. Test connection sends a one-token request to each default deployment (or to your first chat deployments if you have not chosen defaults), through the same path real traffic uses. It is billed to your Azure subscription (cents).
  2. Routing preview shows, for every catalog model, whether Foundry, Vertex AI, or SecureAI will serve it, and with which deployment.
  3. Turn on Route traffic to Azure AI Foundry and click Save.
Models served by your Foundry show the Azure logo in the chat model selector.

​
Data residency

The region requests are processed in depends on the deployment type: SecureAI applies your organization’s residency policy with that region. If you need residency, use Data Zone or regional deployments.

​
What is billed, and to whom

  • LLM usage is billed by Microsoft to your Azure subscription, with your agreement’s discounts and credits.
  • SecureAI shows the cost of those turns as a list-price estimate for analytics only.
  • Turns on Foundry do not deduct points from users unless you turn on Apply user point quotas.

​
Security and compliance

  • Encrypted credential. The client secret is stored encrypted (AES-256-GCM) and is never returned to the browser.
  • Minimal permissions. Cognitive Services OpenAI User to call models and Reader to list deployments, both scoped to the resource — never to the subscription.
  • The path stays protected. Every call goes through SecureAI’s SMLTP gateway (DLP, per-request authorization token, signed receipt). The gateway only connects to <your-resource>.openai.azure.com.
  • Fails closed. If the configuration cannot be read or Entra ID does not issue a token, the request is refused.
  • Audit. Saving, testing, or disconnecting is recorded in the activity log (field names only, never the secret).

​
Troubleshooting

Errors users may see in the chat:

​
Current limitations

  • Claude on Foundry is not supported.
  • Embeddings, re-ranking, OCR, the admin assistant, image generation, and realtime voice stay on SecureAI-managed providers.
  • One Foundry resource per organization.
  • Authentication is by service principal with a client secret (no certificates or managed identity yet).