So you're thinking about the Microsoft AI-102 exam. Maybe you've been building Azure AI solutions and want the credential to prove it. Or maybe you're moving into the AI engineer role and need a study plan that actually works. This article covers everything you need to know about azure ai-102 study guide. This study guide breaks down every domain, the service selection rules that appear in 80%+ of questions, and the specific traps that catch people on exam day. No fluff. I've been working through the practice questions, and these patterns show up again and again. Let's get into it.
Exam at a Glance AI-102 grants the Microsoft Certified: Azure AI Engineer Associate certification. 1008+ practice questions feed into this exam blueprint. You get 40-60 questions, 100 minutes, need 70/100 to pass. | Domain | Weight | Focus | |
|
|
| | Plan and manage an Azure AI solution | 20-25% | Security, deployment, monitoring, Responsible AI | | Implement generative AI solutions | 15-20% | Foundry, RAG, prompts, content safety | | Implement NLP solutions | 15-20% | Language, Speech, Translator | | Implement knowledge mining | 15-20% | Azure AI Search, Document Intelligence | | Implement computer vision | 10-15% | Vision, Custom Vision, OCR | | Implement agentic solutions | 5-10% | Agents, tools, guardrails |
The Service Map You Need Memorized Every scenario question comes down to picking the right service. This table is 70% of the exam: | Requirement | Answer | Not This | |
|
|
| | Extract fields from invoices/receipts | Document Intelligence | Generic OCR, Azure AI Vision | | Ground model on enterprise data | RAG + Azure AI Search | Fine-tuning | | One endpoint for multiple AI services | Multi-service resource | Individual resources | | Least privilege, no keys | Managed identity + RBAC | API keys in config | | Private network, no internet | Private endpoints + VNet | NAT, public endpoints | | Moderate harmful text/images | Content Safety | Azure AI Language | | Sentiment, PII, key phrases | Azure AI Language | Azure AI Speech | | Speech-to-text / text-to-speech | Azure AI Speech | Azure AI Language | | Translate with custom terminology | Custom Translator | Standard Translator | | Image tags, captions, objects | Azure AI Vision | Azure AI Search | | Multi-step agent workflow | Agent with tools | Simple RAG chatbot | | Jailbreak/prompt injection protection | Prompt shields | Object detection |
Plan and Manage (20-25%): Highest Weight This domain bleeds into every other one. The exam pattern is always the same: security first, managed services second, least operational effort third. Security patterns that always win:: Managed identity over API keys (always): RBAC with least privilege over broad contributor roles: Key Vault for secrets over hardcoded values: Private endpoints when any scenario mentions "internal" or "private" Monitoring: Azure Monitor + diagnostic settings + Log Analytics. If the question asks about tracking throttling (429 errors), latency, or failed requests, the answer is always enable diagnostic settings and stream to Log Analytics. Responsible AI: Risk management, oversight, monitoring, safety controls before production. Not "use a bigger model." Throttling (429 responses): Implement retry with exponential backoff and review capacity/quotas. Don't just add Key Vault or change auth.
Generative AI (15-20%) RAG architecture is the single most tested concept. The pipeline: chunk documents โ embed chunks โ store in Azure AI Search vector index โ on query, embed question โ retrieve relevant chunks โ inject into prompt โ generate grounded answer. RAG vs fine-tuning: If the scenario is about giving the model knowledge (company docs, product data), it's RAG. If it's about changing behavior (tone, format), it's fine-tuning. The exam will test this distinction repeatedly. "Fabricates unsupported details" = add grounded prompt with refusal behavior, not increase temperature. Temperature: Reduce for more deterministic output. Increase max tokens for longer responses. Increasing temperature makes things more random, not less. Prompt shields: Protect against jailbreak and prompt injection attacks on public-facing chat apps. Model orchestration: Routing between a fast small model (classification) and a large model (final answers) is orchestration, not semantic ranking.
NLP (15-20%) Three services, constantly confused:: Azure AI Language = text only. Sentiment, PII, key phrases, NER, summarization: Azure AI Speech = audio. Speech-to-text, text-to-speech, keyword recognition, speaker ID: Azure AI Translator = language translation. Custom Translator for domain terminology Keyword recognition (Speech) for wake phrases, not image tagging. Language detection identifies language; sentiment analysis identifies opinion polarity: different features. Multilingual bot with custom Q&A? Route by detected language to language-specific projects, not one giant English-only knowledge base.
Knowledge Mining (15-20%) Document Intelligence = structured field extraction. Use prebuilt models first (invoice, receipt, ID). Use composed models when routing multiple doc types to different extractors through one endpoint. Custom models when prebuilt doesn't fit. Azure AI Search = retrieval engine. Appears everywhere: RAG, knowledge mining, agents, enrichment.: OData filters for structured querying (Rating > 4, sort by date): Custom Web API skills to call external enrichment during indexing: Semantic ranking + vector search + hybrid = best retrieval quality: Indexer can't read private blobs? Check permissions first RAG retrieval irrelevant? Smaller meaningful chunks. Not "disable ranking and filters."
Computer Vision (10-15%) Image classification (categories, no location) vs object detection (bounding boxes needed). The exam loves this distinction. Bounding boxes = region/object positions and sizes in an image. Not sentiment scores. Spatial analysis = people presence and movement from camera video. Not Document Intelligence. Prebuilt Vision for tags/captions without custom training. Custom Vision for your own categories.
Agentic Solutions (5-10%) Choose an agent over a simple RAG chatbot when: multi-step workflow, needs tools, takes actions autonomously. A chatbot answering PDF questions = RAG. A system that reasons over a goal, calls tools, completes a loan document workflow with human approvals = agent. Production safety for multi-agent systems: controlled orchestration, permissions, stopping rules, escalation paths. Not shared admin keys.
