AI on Microsoft Azure: An Overview of Azure AI Services for Developers
A practical developer guide to Azure AI services, mapping each API to real tasks and explaining resources, authentication, pricing models and quota requests.
Azure AI services are Microsoft’s cloud APIs for adding language, speech, vision, document processing and generative AI to applications without training models from scratch. This overview maps each service to the problems it solves, explains how resources, authentication and quotas work, and points out the naming changes that make the documentation confusing for newcomers.
First, a note on names #
Microsoft has renamed this portfolio several times. What began as Cognitive Services became Azure AI services, and the development portal and platform have been branded Azure AI Studio, Azure AI Foundry and, more recently, Microsoft Foundry. Individual services have also changed names: Form Recognizer is now Document Intelligence, and Azure Cognitive Search is now Azure AI Search.
When a tutorial and the portal disagree, the underlying resource type and REST API are usually the same. Search the official documentation using both the old and new name if something seems missing.
Pick a service by the job you need done #
| If you need to… | Look at | Notes |
|---|---|---|
| Generate or summarize text, chat, reason over tools | Azure OpenAI and other models in the Foundry model catalog | Deploy a model, then call it by deployment name |
| Transcribe audio or generate speech | Azure AI Speech | Speech to text, text to speech, speech translation |
| Extract fields from invoices, receipts, IDs or forms | Azure AI Document Intelligence | Prebuilt models plus custom extraction models |
| Analyze sentiment, detect PII, extract key phrases | Azure AI Language | Also includes conversational language understanding and question answering |
| Translate text or documents | Azure AI Translator | Text translation and whole-document translation |
| Tag images, read text in images (OCR) | Azure AI Vision | Image analysis and optical character recognition |
| Build search or RAG over your own content | Azure AI Search | Keyword, vector and hybrid search with ranking |
| Filter harmful text and images | Azure AI Content Safety | Moderation, prompt attack detection, groundedness checks |
Generative AI: Azure OpenAI and the model catalog #
Azure OpenAI provides OpenAI models (such as GPT and embedding models) running in Azure, with Azure networking, identity and compliance controls. The Foundry model catalog adds models from other providers, including open-weight models, some of which are sold directly by Microsoft and some through Azure Marketplace terms.
The workflow differs slightly from calling a public API. You first create a resource in a subscription and Region, then create a deployment of a specific model and version, and your code calls that deployment by name. Deployment types determine where data is processed and how you pay: Global deployments route across Microsoft’s worldwide capacity, Data Zone deployments keep processing within a geographic zone such as the US or EU, and Regional deployments stay in one Region.
Microsoft’s documentation states that prompts and completions sent to Azure OpenAI are not used to train OpenAI models and are not made available to OpenAI. Some data may be stored temporarily for abuse monitoring, and eligible customers can apply to modify that monitoring through the official process.
Prebuilt AI services in practice #
Document Intelligence
Probably the service with the quickest return for business applications. Prebuilt models extract structured fields from invoices, receipts, identity documents, tax forms and general layouts, returning JSON with confidence scores. When prebuilt models do not fit, you can label a small set of your own documents to train a custom extraction model.
Speech
Speech to text supports real-time and batch transcription, with options for speaker diarization and custom vocabulary. Text to speech offers a large catalog of neural voices, controllable with SSML. Custom neural voice, which creates a synthetic version of a specific person’s voice, is a Limited Access feature that requires an application and consent from the voice talent.
Language and Translator
Language covers classic NLP tasks: sentiment, named entities, PII detection and redaction, summarization, and custom text classification. Translator handles text translation across a wide set of languages and can translate whole documents while preserving formatting. For many tasks, a generative model can now do similar work, but these dedicated APIs often cost less, respond faster and behave more predictably for narrow, high-volume jobs.
Vision and Face
Vision provides image tagging, captioning and OCR. The Face service is more tightly controlled: identification and verification features are Limited Access, available only to approved customers for approved use cases, and some capabilities (such as inferring emotional states) have been retired under Microsoft’s Responsible AI Standard.
Resources, endpoints and authentication #
Every service is used through an Azure resource created in a resource group, which gives you an endpoint URL and, by default, two access keys. You can create a single-service resource (for example, just Speech) or a multi-service resource that exposes several services under one endpoint and key, which simplifies setup for prototypes.
- Prefer Microsoft Entra ID over keys. Assign roles such as Cognitive Services User to a managed identity, so your application authenticates without storing secrets. Many teams disable key-based auth entirely once this works.
- If you use keys, store them in Azure Key Vault and rotate them, using the second key to avoid downtime.
- Restrict network access with firewall rules or private endpoints so the resource is not reachable from the whole internet.
- Enable diagnostic logs to a Log Analytics workspace for auditing and troubleshooting.
Where these resources live, and who can manage them, depends on your subscription and resource group structure, which is covered in our guide to Azure tenants, subscriptions and resource groups.
How pricing works #
Pricing differs by service, but the models fall into a few patterns. Check the official Azure pricing pages for current rates in your Region.
- Pay-as-you-go per unit: per 1,000 transactions, per page, per audio hour or per million characters, depending on the service.
- Free (F0) tiers: many services offer a free pricing tier with low monthly limits, suitable for development but not production.
- Commitment tiers: some services offer discounted rates in exchange for a fixed monthly commitment.
- Tokens for generative models: Azure OpenAI standard deployments bill per input and output token; batch deployments are typically cheaper for asynchronous work.
- Provisioned throughput: reserved model processing capacity (measured in provisioned throughput units) for predictable, latency-sensitive workloads, available hourly or with reservations.
Quotas and Limited Access: the official routes #
Generative model deployments are constrained by quota, typically tokens per minute and requests per minute, assigned per model, per deployment type, per Region and per subscription. You divide your quota among deployments. When you need more, request an increase through the quota request process linked from the Foundry portal or open an Azure support request, and describe your workload accurately.
Other services have per-resource rate limits (for example, transactions per second), and exceeding them returns HTTP 429 responses. Build retries with exponential backoff into your client code.
How it compares with other clouds #
Every major cloud now offers managed foundation models alongside task-specific AI APIs. On AWS, the closest counterpart to Azure’s model catalog is Amazon Bedrock, described in our Amazon Bedrock overview. Azure’s distinguishing features are direct access to OpenAI models within Azure’s compliance boundary and close integration with Microsoft 365 and Entra ID. Choose based on where your data and identity already live, not on model headlines alone. The Microsoft Azure overview summarizes the platform more broadly.
Official references: the Azure AI services documentation and the wider Azure documentation, including Microsoft’s Responsible AI transparency notes for each service.
Frequently asked questions #
Do I need a separate resource for each AI service?
No. A multi-service resource gives one endpoint and key for several services, which is convenient for development. Single-service resources give finer control over pricing tiers, networking and access, so production setups often use them.
Why does my Azure OpenAI call fail with a deployment not found error?
Azure OpenAI calls use your deployment name rather than the model name. Check that the deployment exists in the resource whose endpoint you are calling, and that the API version in your request supports that model.
How do I get more tokens-per-minute quota?
Review current allocation in the Foundry portal’s quota view, rebalance across deployments if possible, and submit a quota increase request through the official form or an Azure support request.
Is there a free way to try Azure AI services?
Many services offer a free F0 pricing tier with monthly limits, and new Azure accounts have typically included introductory credit. Check the current offer on Microsoft’s official Azure free account page before relying on it.
