Azure AI Foundry

Learn how to integrate Azure AI Foundry with Prisma AIRS AI Gateway to access a wide range of AI models with enhanced observability and reliability features.

Azure AI Foundry provides a unified platform for enterprise AI operations, model building, and application development. With the AI Gateway, you can seamlessly integrate with various models available on Azure AI Foundry and take advantage of features like observability, fallbacks, and more.

​

Quick Start

cURL

# 1. Add @azure-foundry provider in Model Catalog
# 2. Use it:

curl https://aigw.portkey.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PORTKEY_API_KEY" \
  -d '{
    "model": "DeepSeek-V3-0324",
    "messages": [{"role": "user", "content": "Tell me about cloud computing"}]
  }'

---

​

Add Provider in Model Catalog

To integrate Azure AI Foundry with the AI Gateway, you’ll create a provider in the Model Catalog. This securely stores your Azure AI Foundry credentials, allowing you to use a simple identifier in your code instead of handling sensitive authentication details directly.

OpenAI models on Azure

If you’re specifically looking to use OpenAI models on Azure, you should use Azure OpenAI instead, which is optimized for OpenAI models.

---

​

Understanding Azure AI Foundry Deployments

Azure AI Foundry offers three different ways to deploy models, each with unique endpoints and configurations:

1. AI Services: Azure-managed models accessed through Azure AI Services endpoints
2. Managed: User-managed deployments running on dedicated Azure compute resources
3. Serverless: Seamless, scalable deployment without managing infrastructure

You can learn more about the Azure AI Foundry deployment here.

---

​

Creating Your Azure AI Foundry Provider

Integrate Azure AI Foundry with the AI Gateway to centrally manage your AI models and deployments. This guide walks you through setting up the provider using API key authentication.

​

Prerequisites

Before creating your provider, you’ll need:

• An active Azure AI Foundry account
• Access to your Azure AI Foundry portal
• A deployed model on Azure Foundry

​

Step 1: Navigate to Model Catalog

Go to Model Catalog → Add Provider and select Azure AI Foundry as your provider.

​

Step 2: Configure Provider Details

Fill in the basic information for your provider:

• Name: A descriptive name for this provider (e.g., “Azure AI Production”)
• Short Description: Optional context about this provider’s purpose
• Slug: A unique identifier used in API calls (e.g., “@azure-ai-prod”)

​

Step 3: Set Up Authentication

The AI Gateway supports three authentication methods for Azure AI Foundry. For most use cases, we recommend using the Default (API Key) method.

Default (API Key)

​

Gather Your Azure Credentials

From your Azure AI Foundry portal, you’ll need to collect:

1. Navigate to your model deployment in Azure AI Foundry
2. Click on the deployment to view details
3. Copy the API Key from the authentication section
4. Copy the Target URI - this is your endpoint URL
5. Note the API Version from your deployment URL
6. Azure Deployment Name (Optional): Only required for Managed Services deployments

Azure Managed Entity

For managed Azure deployments:

Required parameters:

• Azure Managed ClientID: Your managed client ID

• Azure Foundry URL: The base endpoint URL for your deployment, formatted according to your deployment type:

  • For AI Services: https://your-resource-name.services.ai.azure.com/models
  • For project-scoped OpenAI v1 deployments: https://your-resource-name.services.ai.azure.com/api/projects/your-project-name/openai/v1
  • For Managed: https://your-model-name.region.inference.ml.azure.com/score
  • For Serverless: https://your-model-name.region.models.ai.azure.com

• Azure API Version: The API version to use (e.g., “2024-05-01-preview”). This is required if you have api version in your deployment url. Examples:

  • If your URL is https://mycompany-ai.westus2.services.ai.azure.com/models?api-version=2024-05-01-preview, the API version is 2024-05-01-preview

• Azure Deployment Name: (Optional) Required only when a single resource contains multiple deployments.

Azure Entra ID

To use this authentication your azure application need to have the role of: conginitive services user. Authentication with Azure Entra ID:

Required parameters:

• Azure Entra ClientID: Your Azure Entra client ID

• Azure Entra Secret: Your client secret

• Azure Entra Tenant ID: Your tenant ID

• Azure Foundry URL: The base endpoint URL for your deployment, formatted according to your deployment type:

  • For AI Services: https://your-resource-name.services.ai.azure.com/models
  • For project-scoped OpenAI v1 deployments: https://your-resource-name.services.ai.azure.com/api/projects/your-project-name/openai/v1
  • For Managed: https://your-model-name.region.inference.ml.azure.com/score
  • For Serverless: https://your-model-name.region.models.ai.azure.com

• Azure API Version: The API version to use (e.g., “2024-05-01-preview”). This is required if you have api version in your deployment url. Examples:

  • If your URL is https://mycompany-ai.westus2.services.ai.azure.com/models?api-version=2024-05-01-preview, the API version is 2024-05-01-preview

• Azure Deployment Name: (Optional) Required only when a single resource contains multiple deployments. Common in Managed deployments.

You can Learn more about these Azure Entra Resources here

For project-scoped OpenAI v1 deployments, enter the full URL ending with /api/projects/<project-name>/openai/v1, such as https://your-resource-name.services.ai.azure.com/api/projects/your-project-name/openai/v1, exactly as shown in Azure AI Foundry. The AI Gateway appends the request route, such as /chat/completions, when sending traffic to the deployment.

​

Adding Multiple Models to Your Azure AI Foundry Provider

You can deploy multiple models through a single Azure AI Foundry provider by using the AI Gateway’s custom models feature.

​

Steps to Add Additional Models

1. Navigate to your Azure AI Foundry provider in Model Catalog
2. Select the Model Provisioning step
3. Click Add Model in the top-right corner

​

Configure Your Model

Enter the following details for your Azure deployment:

Model Slug: Use your Azure Model Deployment name exactly as it appears in Azure AI Foundry

Short Description: Optional description for team reference

Model Type: Select “Custom model”

Base Model: Choose the model that matches your deployment’s API structure (e.g., select gpt-4 for GPT-4 deployments)

This is just for reference. If you can’t find the particular model, you can just choose a similar model.

Custom Pricing: Enable to track costs with your negotiated rates

Once configured, this model will be available alongside others in your provider, allowing you to manage multiple Azure deployments through a single set of credentials.

​

Using Anthropic Models on Azure AI Foundry

Azure AI Foundry supports Anthropic models (Claude) through a slightly different configuration process. Follow these steps to integrate Anthropic models with the AI Gateway.

​

Step 1: Create an Azure Foundry Provider

When creating the provider for Anthropic models, you’ll need to configure the following:

1. Azure API Key: Copy this from your Azure Foundry console

2. Azure Target URI: From your Foundry console, you’ll get a URL like:

   https://resource-name-swedencentral.services.ai.azure.com/anthropic/v1/messages
   

   You need to strip your URL till /anthropic:

   https://resource-name-swedencentral.services.ai.azure.com/anthropic
   

For Anthropic models on Azure Foundry, you don’t need to provide the Azure API Version or Deployment Name fields.

​

Step 2: Configure Workspace Provisioning

After setting up the provider credentials, proceed with the workspace provisioning step as usual.

​

Step 3: Add Your Anthropic Model

In the Model Provisioning step:

1. Click the + Add Model button at the top
2. Configure the model with these details:
   • Model Slug: Enter your deployment name from the Azure Foundry console
   • Base Model: Search for and select your Anthropic model (e.g., claude-opus-4-5-20251101, claude-sonnet-4-5-20250929, claude-haiku-4-5-20251001, claude-opus-4-1-20250805)
3. Save the configuration

​

Making Requests to Anthropic Models

Once configured, you can call your Anthropic model using the Model Slug you saved:

cURL

curl https://aigw.portkey.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PORTKEY_API_KEY" \
  -H "x-portkey-provider: $AZURE_FOUNDRY_ANTHROPIC_PROVIDER" \
  -d '{
    "model": "your-azure-deployment-name",
    "messages": [
      { "role": "user", "content": "Hello, Claude!" }
    ]
  }'

​

Using the /messages Route with Azure Foundry Anthropic Models

Access Anthropic models on Azure AI Foundry through Anthropic’s native /messages endpoint using the AI Gateway’s SDK or Anthropic’s SDK.

The /messages route provides access to Anthropic-native features like extended thinking, prompt caching, and native streaming formats when using Claude models on Azure AI Foundry.

cURL

curl --location 'https://aigw.portkey.ai/v1/messages' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer YOUR_PORTKEY_API_KEY' \
--data-raw '{
    "model": "@your-azure-foundry-anthropic-provider/your-azure-deployment-name",
    "max_tokens": 1024,
    "messages": [
      {
        "role": "user",
        "content": "Hello, Claude!"
      }
    ]
  }'

Anthropic Python SDK

import anthropic

client = anthropic.Anthropic(
    api_key="dummy", # auth happens via the Authorization header
    default_headers={"Authorization": "Bearer YOUR_PORTKEY_API_KEY"},
    base_url="https://aigw.portkey.ai"
)

message = client.messages.create(
    model="@your-azure-foundry-anthropic-provider/your-azure-deployment-name",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "Hello, Claude!"}
    ],
)
print(message.content)

Anthropic TS SDK

import Anthropic from '@anthropic-ai/sdk';

const anthropic = new Anthropic({
  apiKey: 'dummy', // auth happens via the Authorization header
  defaultHeaders: { "Authorization": "Bearer YOUR_PORTKEY_API_KEY" },
  baseURL: "https://aigw.portkey.ai",
});

const msg = await anthropic.messages.create({
  model: "@your-azure-foundry-anthropic-provider/your-azure-deployment-name",
  max_tokens: 1024,
  messages: [{ role: "user", content: "Hello, Claude!" }],
});
console.log(msg);

​

Using the Responses API with Azure AI Foundry

The AI Gateway supports the Responses API on Azure AI Foundry in two modes. Pick the one that matches your Azure deployment.

Mode	Default?	How it works	When to use
Gateway Transformation	Yes	The AI Gateway converts Responses → Chat Completions → Responses around your Azure deployment	Any model deployed on Azure AI Foundry (DeepSeek, Cohere, Llama, Mistral, etc.)
Direct (Proxy)	Opt-in via header	The AI Gateway forwards your Responses payload as-is to Azure’s native /responses endpoint	Azure deployments that natively support /responses (e.g. Azure AI Agents)

In both modes, the AI Gateway logging, tracing, guardrails, caching, retries, and fallbacks continue to work.

​

Mode 1 - Gateway Transformation (Default)

This is the default behaviour and requires no extra configuration. The AI Gateway accepts the request at /v1/responses, adapts the payload to Chat Completions, calls Azure AI Foundry’s /chat/completions endpoint, and translates the response back to the Responses API format before returning it to you.

Because the call is made over Chat Completions on Azure’s side, every model on your Azure AI Foundry provider works — even ones that don’t natively expose a Responses endpoint.

cURL

curl https://aigw.portkey.ai/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PORTKEY_API_KEY" \
  -d '{
    "model": "@AZURE_FOUNDRY_PROVIDER/DeepSeek-V3-0324",
    "input": "Explain the difference between AKS and ACI in one paragraph."
  }'

Features that rely on Azure storing response state — previous_response_id, store, retrieve/delete on /v1/responses/:id, and built-in tools like web_search — are not available in this mode. Use multi-turn input arrays to carry conversation history. See Open Responses → Native-Only Features.

​

Mode 2 - Direct Responses (Proxy)

When your Azure AI Foundry deployment exposes a native /responses endpoint (for example, Azure AI Agents), set the x-portkey-provider-responses-proxy header to true. The AI Gateway will skip the Chat Completions transformation and forward the request body exactly as-is to Azure’s native Responses endpoint.

This is required to use Azure-specific parameters such as agent_reference, which targets a deployed Azure AI Agent.

cURL

curl https://aigw.portkey.ai/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PORTKEY_API_KEY" \
  -H "x-portkey-provider-responses-proxy: true" \
  -d '{
    "model": "@AZURE_FOUNDRY_PROVIDER/gpt-4.1",
    "input": "Summarise the latest Q3 sales numbers for the EMEA region.",
    "agent_reference": "asst_abc123"
  }'

Direct mode forwards the payload verbatim. Your Azure deployment must accept the Responses API schema — if it doesn’t, requests will fail. Use Gateway Transformation (Mode 1) for everything else.

​

Choosing the right mode

• Use Gateway Transformation for general-purpose Responses API calls against any Azure-deployed model.
• Use Direct (Proxy) when calling Azure AI Agents or any Azure deployment that exposes the native Responses endpoint and you need Azure-specific parameters (agent_reference, server-side response storage, etc.).

For a deep dive into the Responses API itself — streaming, tool calling, reasoning, vision, structured output — see the Open Responses guide.

​

Sample Request

Once you’ve created your provider, you can start making requests to Azure AI Foundry models through the AI Gateway.

cURL

curl https://aigw.portkey.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PORTKEY_API_KEY" \
  -H "x-portkey-provider: $AZURE_FOUNDRY_PROVIDER" \
  -d '{
    "model": "DeepSeek-V3-0324",
    "messages": [
      { "role": "user", "content": "Tell me about cloud computing" }
    ]
  }'

NodeJS

Install the OpenAI SDK with npm

npm install openai

Python

Install the OpenAI SDK with pip

pip install openai

​

Advanced Features

​

Function Calling

Azure AI Foundry supports function calling (tool calling) for compatible models. Here’s how to implement it with the AI Gateway:

​

Vision Capabilities

Process images alongside text using Azure AI Foundry’s vision capabilities:

​

Structured Outputs

Get consistent, parseable responses in specific formats:

---

​

AI Gateway Features

The AI Gateway provides advanced gateway features for Azure AI Foundry deployments:

​

Fallbacks

Create fallback configurations to ensure reliability when working with Azure AI Foundry models:

{
  "strategy": {
    "mode": "fallback"
  },
  "targets": [
    {
      "provider": "@azure-foundry-prod",
      "override_params": {
        "model": "DeepSeek-V3-0324"
      }
    },
    {
      "provider": "@openai-prod",
      "override_params": {
        "model": "gpt-4o"
      }
    }
  ]
}

​

Load Balancing

Distribute requests across multiple models for optimal performance:

{
  "strategy": {
    "mode": "loadbalance"
  },
  "targets": [
    {
      "provider": "@azure-foundry-prod-1",
      "override_params": {
        "model": "DeepSeek-V3-0324"
      },
      "weight": 0.7
    },
    {
      "provider": "@azure-foundry-prod-2",
      "override_params": {
        "model": "cohere-command-a"
      },
      "weight": 0.3
    }
  ]
}

​

Conditional Routing

Route requests based on specific conditions like user type or content requirements:

{
  "strategy": {
    "mode": "conditional",
    "conditions": [
      {
        "query": { "metadata.user_type": { "$eq": "premium" } },
        "then": "high-performance-model"
      },
      {
        "query": { "metadata.content_type": { "$eq": "code" } },
        "then": "code-specialized-model"
      }
    ],
    "default": "standard-model"
  },
  "targets": [
    {
      "name": "high-performance-model",
      "provider": "@azure-foundry-prod-1",
      "override_params": {
        "model": "Llama-4-Scout-17B-16E"
      }
    },
    {
      "name": "code-specialized-model",
      "provider": "@azure-foundry-prod-2",
      "override_params": {
        "model": "DeepSeek-V3-0324"
      }
    },
    {
      "name": "standard-model",
      "provider": "@azure-foundry-prod-3",
      "override_params": {
        "model": "cohere-command-a"
      }
    }
  ]
}

---

​

Rerank

Azure AI Foundry supports reranking through Cohere models deployed on the platform. Use the AI Gateway unified /rerank endpoint with the cohere. model prefix:

cURL

curl https://aigw.portkey.ai/v1/rerank \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PORTKEY_API_KEY" \
  -d '{
    "model": "@azure-foundry/cohere.Cohere-rerank-v4.0-pro",
    "query": "What is deep learning?",
    "documents": [
      "Deep learning is a subset of machine learning",
      "The weather is sunny today",
      "Neural networks have multiple layers"
    ]
  }'

The cohere. prefix in the model name is automatically stripped before forwarding to the provider.

---

​

OCR (Document Processing)

Azure AI Foundry supports Mistral OCR models for extracting text and structured content from PDFs and images. Use the /v1/ocr endpoint:

cURL

curl https://aigw.portkey.ai/v1/ocr \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PORTKEY_API_KEY" \
  -d '{
    "model": "@azure-foundry/mistral-ocr-4-0",
    "document": {
      "type": "document_url",
      "document_url": "data:application/pdf;base64,<BASE64_PDF>"
    },
    "include_image_base64": true
  }'

Supported Models: mistral-ocr-4-0, mistral-document-ai-2505, mistral-document-ai-2512

Azure AI Foundry does not support external URLs for OCR. Documents must be sent as base64-encoded data URIs (e.g. data:application/pdf;base64,...).

The native /v1/ocr endpoint requires gateway version 2.18.0 or higher for hybrid (self-hosted) deployments.

---

​

Decisions (Microsoft-Decision-1)

Microsoft-Decision-1 is a decision model available in Microsoft Foundry. It does not generate text. It reads a piece of text or JSON, answers typed questions about it, and returns probabilities that your application can threshold or rank. Use it for classification, routing, ranking, grading, and yes/no checks.

The AI Gateway routes these calls through /v1/decisions, a separate endpoint from chat completions. You still get the gateway’s full feature set: authentication, routing, retries, circuit breaking, caching, logging, and cost tracking. See the Decisions API for the full request and response reference.

/v1/decisions does not support streaming. Only before-request guardrails apply, scanning the content to judge only (state).

This section covers how to call Microsoft-Decision-1 through the AI Gateway. For model details, deployment steps in Azure, regional availability, and responsible-use guidance, see Microsoft’s official guide: Deploy and use Microsoft-Decision-1 in Microsoft Foundry.

​

Model at a glance

These values come from Microsoft’s model overview.

Model name	Model version	Deployment type	API type
Microsoft-Decision-1	1	DataZoneStandard in selected regions, GlobalStandard	Decision

​

Request model vs response model

This is the most common source of errors with this model, so read it before you send a request.

Field	Value	Example
model in your request	The deployment name you chose in Azure AI Foundry	@azure-foundry/decision-1
model in the response	The underlying model	microsoft-decision-1

Do not hardcode microsoft-decision-1 as the request model. Azure looks up a deployment with that exact name. If your deployment has a different name, the request fails with The API deployment microsoft-decision-1 does not exist. Use the deployment name you see in Foundry. Microsoft describes this in the Call the decision API section of its guide.

For example, if you deploy the model with the deployment name decision-1, send "model": "@azure-foundry/decision-1". If you name the deployment pi-decision-1, send "model": "@azure-foundry/pi-decision-1". In both cases the response reports "model": "microsoft-decision-1".

The AI Gateway logs and prices the call using the response model, so cost tracking works the same no matter what you name your deployment.

​

Setup

1. Deploy Microsoft-Decision-1 in your Foundry project by following Microsoft’s deployment steps. From the deployment details page, copy the resource endpoint, the deployment name, and the API key.
2. Create your Azure AI Foundry provider by following Creating Your Azure AI Foundry Provider above. Use the URL as shown in Foundry for your resource, for example https://your-resource-name.services.ai.azure.com/models.
3. If you manage models in Model Provisioning, add your deployment name as the model slug (for example decision-1), exactly as it appears in Foundry. See Adding Multiple Models.

For decisions requests, The AI Gateway sends the call to the resource root at /providers/microsoft/v1/systemone. Any path in the saved Foundry URL, including a /deployments/<name> suffix, is dropped. No api-version is sent. The deployment is selected by the model field in the request body.

Decisions requests work with the same authentication methods as the rest of this integration: API key, Azure Entra ID, and Azure Managed Identity. With an API key, The AI Gateway sends it in the api-key header, which is what the Microsoft decision endpoint expects.

​

Send a decisions request

Microsoft-Decision-1 uses the Jev request shape: state for the content to judge, and a questions map. Set model to @provider-slug/deployment-name. The provider holds the Azure credentials, so no provider headers are needed.

cURL

curl --request POST \
  --url https://aigw.portkey.ai/v1/decisions \
  --header 'Content-Type: application/json' \
  --header "Authorization: Bearer $PORTKEY_API_KEY" \
  --data '{
  "model": "@azure-foundry/decision-1",
  "state": {
    "current_ask": "hello there, how is your day going?"
  },
  "questions": {
    "greeting": {
      "type": "choice",
      "instructions": "Is this a greeting?",
      "criteria": {
        "yes": "a greeting",
        "no": "anything else"
      }
    }
  }
}'

Response (values are illustrative):

{
  "model": "microsoft-decision-1",
  "answers": {
    "greeting": {
      "type": "choice",
      "choice": "yes",
      "probabilities": { "no": 0.0067, "yes": 0.9933 },
      "confidence": 0.9866
    }
  },
  "usage": { "input_tokens": 46, "output_tokens": 1 },
  "provider": "azure-ai"
}

The OpenAI request shape (input and a questions array) also works. See Request Shapes.

​

Question types

Each entry in questions sets a type, which determines the shape of its answer. The state can be text or JSON.

Type	Question it answers	Answer shape
noul	Is this true?	{ noul: <0 to 1 probability> }
choice	Which option is it?	{ choice, probabilities, confidence }
score	How much?	{ score, legend, probabilities, confidence }

The response answers object is keyed by the question names in your request.

​

noul: gates and filters

Use for yes/no conditions, such as whether a reply promises a refund or a document contains a prompt injection. criteria is optional. Define the true and false conditions to make borderline questions more precise.

{
  "model": "@azure-foundry/decision-1",
  "state": "I refunded your $200. You don't need to contact billing.",
  "questions": {
    "promises_refund": {
      "type": "noul",
      "instructions": "Does this reply promise the customer a refund?",
      "criteria": {
        "true": "Commits to returning money",
        "false": "Makes no commitment about money"
      }
    }
  }
}

​

choice: classification and routing

Use to pick exactly one label from a fixed set, such as routing tickets or classifying intent. Describe each option clearly. The response includes a probability for every option, so a low top probability is a good signal to escalate to a person or another model.

{
  "model": "@azure-foundry/decision-1",
  "state": "The API returns 500 on every call.",
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this ticket?",
      "criteria": {
        "billing": "Charges, invoices, and refunds",
        "engineering": "Bugs, errors, and outages",
        "support": "How-to and account questions"
      }
    }
  }
}

​

score: ranking and grading

Use to place an item on an ordered scale, such as incident severity. List the levels from lowest to highest. The returned score is the probability-weighted average of the level indexes, so a four-level scale runs from 0 through 3 and the result can fall between levels. Prefer scores for relative ordering and thresholds, not as absolute ratings.

{
  "model": "@azure-foundry/decision-1",
  "state": "Checkout is down for all EU customers since 09:00.",
  "questions": {
    "severity": {
      "type": "score",
      "instructions": "How severe is this incident?",
      "criteria": ["Cosmetic", "Minor", "Major", "Critical"]
    }
  }
}

​

Combine questions in one request

Questions that share the same state can go in one request. This reduces request count and keeps the context consistent.

{
  "model": "@azure-foundry/decision-1",
  "state": {
    "message": "Checkout is down again. This is my third request.",
    "customer_contact_count": 3
  },
  "questions": {
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this request?",
      "criteria": {
        "billing": "Charges, invoices, and refunds",
        "engineering": "Bugs, errors, and outages",
        "support": "How-to and account questions"
      }
    },
    "severity": {
      "type": "score",
      "instructions": "How severe is this issue?",
      "criteria": ["Cosmetic", "Minor", "Major", "Critical"]
    },
    "repeat_contact": {
      "type": "noul",
      "instructions": "Has the customer contacted support before?"
    }
  }
}

​

Pricing and usage

Usage is reported in the response as usage.input_tokens and usage.output_tokens. The pricing entry for microsoft-decision-1 charges for input tokens only. Output tokens, which hold the typed judgment itself, are priced at zero.

​

Decisions troubleshooting

Status	Likely cause	What to do
400	Invalid question type or shape	Check state, questions, type, and criteria
401	Missing, invalid, or expired credential	Check the API key or Entra credentials on the provider
403	The identity has no access to the deployment	Check role assignments and deployment access in Azure
404	Wrong endpoint, or the deployment name does not exist	Confirm the model value matches your deployment name exactly. See Request model vs response model
422	Unsupported deployment name or question definition	Check the model value and each question definition
429	The deployment rate limit was exceeded	Retry with backoff, or request more quota in Azure. You can also add retries or fallbacks in a gateway config

Upstream errors use the standard error shape, with the message prefixed by azure-ai:

{
  "error": {
    "message": "azure-ai error: The API deployment microsoft-decision-1 does not exist. If you created the deployment within the last 5 minutes, please wait a moment and try again.",
    "type": null,
    "param": null,
    "code": null
  },
  "provider": "azure-ai"
}

​

Limitations and responsible use

From Microsoft’s guidance for this model:

• No explanations. The model returns probabilities, not a written rationale.
• Wording sensitivity. Scores can change with how you phrase or order questions and options.
• Bias. The model might reflect biases from its base model and training data. Do not use its scores as the sole basis for decisions about individuals.
• Safety filtering. If you use it as a safety filter, it can miss subtle harmful content or flag benign content.
• Validate on your data. Test thresholds against labeled examples from your own workload before you automate a gate.

See Known limitations and best practices in Microsoft’s guide for the full list.

---

​

Next Steps

Azure OpenAI

Use Azure OpenAI for OpenAI-specific models

Add Metadata

Add custom metadata to your Azure AI Foundry requests

Gateway Configs

Configure advanced gateway features

Request Tracing

Trace your Azure AI Foundry requests

Setup Fallbacks

Create fallback configurations between providers