AI-102: Exam Guide
How the Exam Wants You to Think
The AI-102 exam is for Azure AI Engineers — developers who build and deploy AI solutions using Microsoft's services and SDKs. It values implementation knowledge: which service to use, how to configure it, and how to integrate it into applications.
Answer Philosophy
- Choose the managed service over DIY — Microsoft wants you to use Azure AI services rather than build from scratch. RAG over custom training; prebuilt models over custom when possible.
- Foundry is the platform — Everything lives in Microsoft AI Foundry. Hubs manage shared infrastructure; Projects are your workspace. When in doubt, the answer involves Foundry.
- SDK over REST — The exam prefers
DefaultAzureCredential()and the Foundry SDK over raw REST calls or hardcoded keys. - Async patterns for heavy operations — OCR Read API, Document Translation, and batch operations all follow the
202 → Operation-Location → GETasync pattern.
Keyword Detection Table
| If you see... | Look for this in the answer... |
|---|---|
| "avoid hardcoded keys" / "keyless auth" | Managed Identity + DefaultAzureCredential() |
| "predictable latency" / "high throughput" | Provisioned Throughput (PTU) |
| "data residency" / "offline" / "edge" | Docker container deployment |
| "build, test, deploy AI apps" | Microsoft AI Foundry Project |
| "shared compute, connections, security" | Microsoft AI Foundry Hub |
| "ground the model in your own data" | RAG (On Your Data / AI Search) |
| "specific tone, format, or rare domain" | Fine-tuning |
| "visual LLM workflow" / "evaluate prompts" | Prompt Flow |
| "block prompt injection attacks" | Prompt Shields |
| "autonomous multi-step task" | AI Agent Service |
| "agent uses Python to solve math" | Code Interpreter tool |
| "agent searches uploaded documents" | File Search tool |
| "multiple agents collaborating" | Multi-agent orchestration |
| "handwritten text extraction" | Read API (OCR 4.0) |
| "locate objects with bounding boxes" | Custom Vision — Object Detection |
| "1:1 face comparison" | Face Verification |
| "1:N face comparison against known people" | Face Identification + PersonGroup |
| "recognize spoken intent / wake word" | Speech SDK — Intent Recognition / Keyword |
| "translate entire Word/PDF, preserve layout" | Document Translation (async) |
| "utterance → intent + entities" | CLU (Conversational Language Understanding) |
| "multi-turn Q&A from documents" | Custom Question Answering |
| "text from images/tables in complex docs" | Content Understanding / Document Intelligence |
| "enrich documents with AI before indexing" | Azure AI Search Skillset |
| "Power BI analytics from enriched docs" | Knowledge Store — Table Projections |
| "custom extract logic in skillset" | Custom Skill (Azure Function) |
| "keyword + vector combined search" | Hybrid Search |
| "re-rank to surface single best answer" | Semantic Ranking |
Exam Traps
Watch out for these common mistakes!
RAG vs Fine-tuning: RAG = runtime context injection (fast, cheap, updatable). Fine-tuning = baked-in knowledge (expensive, slow, better for tone/format). The exam frequently asks "which approach?" — if the data changes often → RAG. If the style/format must be rigid → Fine-tuning.
Hub vs Project: Hub = shared infrastructure (compute, connections, role assignments). Project = your workspace inside a hub. A question about "setting up shared compute for multiple teams" → Hub. "Building a specific chatbot" → Project.
PTU vs Standard: Standard = pay-per-token, variable latency. PTU = reserved capacity, consistent latency, higher fixed cost. Exam uses "predictable latency" or "guaranteed throughput" as the signal for PTU.
Content Understanding vs Document Intelligence: Content Understanding (new) handles multimodal pipelines (images, video, audio + docs) with AI summarization. Document Intelligence = forms and structured extraction with prebuilt/custom models. Both extract from documents but different use cases.
Custom Skill interface: A Custom Skill must follow the exact input/output schema expected by Azure AI Search. The skill receives a
valuesarray and must return avaluesarray with the same record keys. Forgetting this schema is a common mistake.Async OCR pattern: The Read API returns
202 Acceptedwith anOperation-Locationheader. You must thenGETthat URL and poll untilstatus: succeeded. Many candidates try to use the response from the initial POST.PersonGroup vs FaceList: PersonGroup (and LargePersonGroup) is for Identification (1:N — "who is this?"). FaceList (and LargeFaceList) is for Find-Similar (1:N — "find faces similar to this one"). The training step is required for PersonGroup, not FaceList.
Semantic Ranking vs Vector Search: Vector search finds semantically similar documents using embeddings. Semantic ranking re-ranks already-retrieved results using an LLM to surface the single best answer. They are not the same — Semantic Ranking is a post-retrieval step.
Content filters scope: Azure OpenAI content filters apply to the model's inputs AND outputs. Azure AI Content Safety is a separate standalone service for user-generated content moderation. Don't confuse the two.
Decision Quick Reference
"Which generative AI approach?"
Data changes often, reduce hallucinations → RAG (On Your Data)
Specific tone, format, domain jargon → Fine-tuning
Visual workflow, test prompt variants → Prompt Flow
Autonomous multi-step reasoning → AI Agent Service"Which vision service?"
General image analysis, OCR, tagging → Image Analysis 4.0 (Azure Vision)
Custom categories / bounding boxes → Custom Vision
Video insights (faces, brands, topics) → Video Indexer
Real-time movement in video feed → Spatial Analysis
Face verification / identification → Face API"Which NLP service?"
Analyze existing text (sentiment, NER) → Language Service
Understand spoken intent / commands → CLU + Speech SDK
Q&A from documents → Custom Question Answering
Translate text or documents → Translator Service"Which search / extraction approach?"
Search structured and unstructured data → Azure AI Search
Extract fields from forms and invoices → Document Intelligence
AI-enriched extraction pipeline → AI Search Skillset
New multimodal document pipeline → Content Understanding"Which authentication?"
Production app, avoid key rotation → Managed Identity (DefaultAzureCredential)
Simple testing / scripts → Subscription key
Cross-service access, audit trail → RBAC roles2025 Exam Domain Weights
| Domain | Weight |
|---|---|
| 1: Plan and manage an Azure AI solution | 20-25% |
| 2: Implement generative AI solutions | 15-20% |
| 3: Implement an agentic solution | 5-10% |
| 4: Implement computer vision solutions | 10-15% |
| 5: Implement NLP solutions | 15-20% |
| 6: Implement knowledge mining and information extraction | 15-20% |
High-Value Focus
Domains 1, 5, and 6 each carry 15-20%+ weight and together represent over half the exam. Domain 3 (Agentic) is new and lightly weighted — know the concepts but do not over-invest.
Final Strategy
- Know your async patterns cold — Read API, Document Translation, and batch operations all follow
202 → Operation-Location → GET. This pattern comes up repeatedly. - "Foundry" is the answer to "where" — Hubs, projects, Prompt Flow, model catalog, deployments. When a question is about the platform or portal → AI Foundry.
- Eliminate third-party answers — If a choice involves building something from scratch when an Azure service exists, it is almost certainly wrong.
- D1 + D5 + D6 = 50-65% of the exam — Prioritise Plan & Manage, NLP, and Knowledge Mining for maximum return on study time.