An Azure AI engineer's first job is picking the right tool. Azure offers a family of prebuilt AI services, each trained by Microsoft for one kind of problem, plus Azure OpenAI and the wider model catalog for generative AI. Many AI-102 questions describe a business need in a sentence or two and ask which service fits, so you need a clear mental map of what each one does and, just as important, what it does not do.
Azure OpenAI (reached through Azure AI Foundry) hosts large language models for chat, summarization, code, reasoning and embeddings, plus image generation models. Use it when the output is generated content or when you need flexible reasoning over text. Azure AI Search indexes your content for keyword, vector and hybrid search, and is the usual retrieval layer for retrieval augmented generation (RAG). Azure AI Vision analyzes images: captions, tags, objects, people and text through its Read (optical character recognition, OCR) feature. Custom Vision trains your own image classifiers and object detectors, and Azure AI Face detects faces and, with approval, identifies people.
Azure AI Language covers text analytics (sentiment, key phrases, entities, personally identifiable information detection), conversational language understanding (CLU), custom question answering and custom text classification. Azure AI Speech converts speech to text and text to speech, translates speech and recognizes wake words. Azure AI Translator translates text and whole documents. Azure AI Document Intelligence extracts text, tables and named fields from forms such as invoices, receipts and ID documents, and trains custom extraction models. Azure AI Content Understanding is a newer service that extracts fields from documents, images, audio and video with one schema.
The exam likes near misses. Reading text from a photo of a sign is Vision Read, but pulling the total and due date from an invoice is Document Intelligence, because you want named fields, not just lines of text. Detecting that a review is negative is Language sentiment analysis, but deciding whether a post contains hate or violence is Azure AI Content Safety. Translating a chat message is Translator; translating a live speaker is Speech translation. When a prebuilt capability covers the need, prefer it over training a custom model or writing a prompt, because it is cheaper, faster and already evaluated.
Finally, remember the generative option. A large language model can often do sentiment or extraction too, but a purpose-built service is usually cheaper, more predictable and easier to evaluate. Choose Azure OpenAI when the task needs open-ended generation, reasoning or conversation.
Key terms
- Azure AI services
- Microsoft's family of prebuilt AI APIs for vision, language, speech, translation, documents and content safety, called over REST or SDKs.
- Azure OpenAI
- Azure-hosted OpenAI models for chat, reasoning, embeddings and image generation, deployed and managed through Azure AI Foundry.
- Document Intelligence
- A service that extracts text, structure and named fields from documents using prebuilt or custom models.
- Content Safety
- A service that detects harmful text and images across hate, sexual, violence and self-harm categories.
An insurance company wants to process claim forms, let agents search past claims in plain language and read summaries aloud to visually impaired staff. The engineer uses Document Intelligence to extract fields from the forms, Azure AI Search with vector and keyword indexing for retrieval, Azure OpenAI to summarize claims, and Speech text to speech to read the summaries.
Check yourself
A company wants to know whether customer reviews are positive or negative. Which service is the most direct fit?
Azure AI Language sentiment analysis. It is purpose-built, returns scores per sentence and is cheaper and more predictable than prompting a large language model.
Which service would you use to find the text on a street sign in a photo, and which to get the total from a receipt?
Azure AI Vision's Read (OCR) feature for the sign; Document Intelligence's prebuilt receipt model for the receipt total, because it returns a named field.