1.1 Azure AI Services Overview and Selection

Quick Answer: Azure AI services include Azure AI Vision, Azure AI Language, Azure AI Speech, Azure OpenAI Service, Azure AI Search, Azure AI Document Intelligence, Azure AI Content Safety, and Azure AI Translator. Deploy them as a single multi-service resource (one endpoint, one key) or as dedicated single-service resources. Azure AI Foundry is the portal where you build, evaluate, and deploy.

Why service selection dominates Domain 1

The AI-102 exam (Designing and Implementing a Microsoft Azure AI Solution) was the single requirement for the Microsoft Certified: Azure AI Engineer Associate credential. It ran 100 minutes, cost roughly 165 USD, and the passing score was 700 on a 1-1,000 scale. Domain 1 was 20-25% of the score, and within it, service-selection items appear constantly: a short business scenario, then "which service?" The trick is to map the verb in the scenario to exactly one service. Memorizing the table below earns easy points.

Service catalog and the verb that triggers it

Multi-service vs. single-service resources

Azure gives you two provisioning shapes, and the exam loves to contrast them.

Multi-service resource, now the Microsoft Foundry resource (--kind AIServices, formerly the Azure AI services resource): one resource exposes one endpoint and one pair of keys that reach Azure OpenAI, Vision, Language, Speech, Translator, Document Intelligence, and Content Safety together. Billing is consolidated on a single line item. It is ideal for prototypes and for apps that touch several services. Its limitation is that you cannot apply per-service network or RBAC isolation. The older multi-service kind CognitiveServices still exists, but Microsoft's quickstart says to create kind AIServices.

Single-service resource (--kind ComputerVision, TextAnalytics, SpeechServices, OpenAI, etc.): a dedicated endpoint and keys for one capability. You get granular billing, independent quotas, per-service RBAC, and the ability to lock down each service separately — the production-grade choice.

# Multi-service: one endpoint covers many capabilities
https://my-ai-services.cognitiveservices.azure.com/

# Single-service: a distinct endpoint per capability
https://my-vision.cognitiveservices.azure.com/
https://my-language.cognitiveservices.azure.com/
https://my-openai.openai.azure.com/   # OpenAI uses its own *.openai.azure.com host

On the Exam: "single endpoint for multiple cognitive capabilities" => multi-service resource. "granular billing," "per-service access control," or "separate quotas" => single-service. Any mention of GPT, embeddings, or image generation (gpt-image; older items say DALL-E) => an Azure OpenAI model deployment, either in a Microsoft Foundry resource (kind AIServices) or in a standalone Azure OpenAI resource (kind OpenAI).

Azure AI Foundry: hubs, projects, and the portal

Azure AI Foundry (formerly Azure AI Studio) is the build-evaluate-deploy portal. It surfaces a model catalog (OpenAI, Meta Llama, Mistral, Cohere, Microsoft Phi), prompt flow for visual pipeline design (retiring April 20, 2027; Microsoft points new work to Microsoft Agent Framework), evaluations for groundedness/quality/safety, content safety configuration, deployments with versioning, and an agent builder. Foundry organizes work into two containers:

• Hub — a top-level container holding shared compute, storage, key vault, and connections that multiple projects reuse. Think of it as a resource group purpose-built for AI.
• Project — a workspace inside a hub for one solution, with its own deployments, data, indexes, and evaluation runs.

Selection decision tree

• Extract sentiment / entities / PII / key phrases -> Azure AI Language
• Translate text or whole documents -> Azure AI Translator
• Pull fields from invoices, receipts, ID cards -> Document Intelligence
• Generate, summarize, or chat over text -> Azure OpenAI Service
• Classify your own labeled images -> Custom Vision; read text from images -> Vision (Read)
• Convert speech to text or text to speech -> Azure AI Speech
• Search + ground a chatbot on enterprise data -> Azure AI Search + Azure OpenAI (RAG)
• Moderate harmful text/images -> Content Safety

Common selection traps

The exam deliberately blurs services with overlapping surface area, so learn the boundaries. Document Intelligence vs. Vision Read: both pull text from images, but if the scenario names invoices, receipts, ID cards, W-2s, or structured field extraction, the answer is Document Intelligence; plain unstructured text extraction from a photo is Vision (Read). Custom Vision vs. Vision Image Analysis: Image Analysis returns Microsoft's prebuilt tags and captions, while Custom Vision is the only choice when the scenario says you have your own labeled training images and need domain-specific classes. Azure AI Language CLU vs. Azure OpenAI: Conversational Language Understanding maps utterances to a fixed, trained set of intents and entities (deterministic, no hallucination); Azure OpenAI generates free-form text and is the answer when the scenario wants open-ended summaries, drafting, or chat. Azure AI Search vs. a database: Search adds full-text, vector, hybrid retrieval, and semantic ranking over your content, and it is the retrieval half of every RAG question.

A second recurring trap is conflating Azure AI Content Safety (the standalone moderation service with Prompt Shields and groundedness) with the built-in content filters inside Azure OpenAI. Standalone Content Safety is used to moderate arbitrary user content or images anywhere in a pipeline; the Azure OpenAI filters apply specifically to that deployment's prompts and completions. When a scenario moderates user-generated content that never touches a GPT model, the answer is the standalone Content Safety service.

On the Exam: Read the scenario for a single decisive noun ("invoice," "my own images," "fixed set of intents," "generate a summary"). That noun usually eliminates three of the four options before you weigh anything else, which is why the verb-plus-noun mapping above is worth memorizing cold.