Microsoft AI-102 Azure AI Engineer Associate: The Cert That Proves You Can Build AI Solutions at Scale

Domain 1 — Plan and Manage an Azure AI Solution (~15–20%)

The foundation domain: selecting the right Azure AI service for a given business requirement, provisioning resources, configuring authentication, and monitoring deployed solutions. The exam tests decision-making, not just configuration.

• Service selection: Azure AI Services multi-service resource vs single-service resource vs Azure Machine Learning. Multi-service consolidates billing and simplifies key management for teams using multiple cognitive services; single-service provides isolated rate limits and fine-grained cost attribution per workload.
• Authentication: managed identity (system-assigned vs user-assigned) preferred over API keys for Azure-to-Azure access. API keys for external client applications with short rotation cycles. Microsoft Entra ID tokens for enterprise single sign-on scenarios.
• Containers: Azure AI Services containerised endpoints for air-gapped or on-premises deployments. Know which services support container deployment (Language, Speech, Vision, Form Recognizer) and the billing model (billed on Azure even when running on-premises via container).
• Monitoring: Azure Monitor metrics for service availability and latency; Application Insights for usage telemetry from custom AI applications; diagnostic logs to Log Analytics for audit trails and anomaly detection.
• Responsible AI principles: fairness, reliability, privacy, inclusiveness, transparency, accountability. Microsoft’s Content Safety service for detecting harmful content across text and image inputs — tested in the context of content moderation pipelines and LLM output filtering.

Domain 2 — Implement Decision Support Solutions (~10–15%)

Azure AI Anomaly Detector: time-series analysis for detecting outliers in metric streams. Two modes — batch (full historical series, best for offline analysis) and streaming (real-time, lower latency, fewer data points). Know the sensitivity parameter: higher sensitivity catches more anomalies but increases false positives.

Azure AI Content Safety: detects hate speech, violence, sexual content, and self-harm across text and images. Severity levels 0–6; the exam tests setting custom thresholds per category based on application context (a children’s platform requires stricter violence thresholds than a professional content moderation tool). Groundedness detection in LLM outputs: verifying that model responses are grounded in the supplied context, not hallucinated.

• Azure Metrics Advisor: enterprise-scale anomaly detection service over multiple time-series streams with root-cause correlation across metrics. Distinct from Anomaly Detector: Metrics Advisor includes a managed UI for business users; Anomaly Detector is an API-first service for developer integration.
• The exam tests selecting Anomaly Detector for developer-controlled API scenarios and Metrics Advisor for business metric monitoring with feedback loops and root-cause graphs.

Domain 3 — Implement Computer Vision Solutions (~15–20%)

The core decision tree candidates must master:

• Azure AI Vision (Image Analysis): pre-built models for general-purpose image analysis — object detection, image description, OCR (Read API), face detection, brand detection, and scene classification. No training data required; use when off-the-shelf accuracy on standard image types is sufficient.
• Custom Vision: transfer learning on custom image datasets for classification (multi-class, multi-label) and object detection when pre-built models underperform on domain-specific images (medical imaging, industrial inspection). Exports to ONNX, TensorFlow, and CoreML for edge deployment. Choose Custom Vision when you have labelled training data and need domain-specific accuracy.
• Azure AI Video Indexer: deep analysis of video files — speaker diarisation, transcript, keyframe extraction, brand and face recognition, sentiment tracking. Used when video is the input format rather than still images.
• Azure AI Face: face detection (bounding boxes, attributes), face verification (1:1 identity check), face identification (1:N matching against a known person group), and liveness detection. Limited Access policy applies — some Face capabilities (identification, verification, emotion) require approval justification from Microsoft.
• Spatial analysis: real-time people-counting and zone-entry detection from RTSP video streams. Requires Azure Stack Edge or Kubernetes deployment for on-premises camera feeds.

The exam repeatedly tests OCR service selection: the Read API in Azure AI Vision for general-purpose printed and handwritten text extraction; Azure AI Document Intelligence for structured document layouts with field extraction (forms, invoices, receipts). The distinction is whether the output is raw text (Read API) or labelled key-value pairs tied to a document schema (Document Intelligence).

Domain 4 — Implement Natural Language Processing Solutions (~15–20%)

Azure AI Language is the consolidated service (formerly Text Analytics, LUIS, QnA Maker). The exam tests choosing the right Language feature for a given scenario:

• Sentiment analysis and opinion mining: document-level and sentence-level sentiment (positive, negative, neutral, mixed) with aspect-based opinion mining for extracting fine-grained sentiment about specific entities (“the battery life is excellent” vs “the camera is disappointing”).
• Named Entity Recognition (NER) and PII detection: extracting entities (person, location, organisation, date, quantity) and personally identifiable information from text. PII detection supports redaction for GDPR compliance pipelines.
• Key phrase extraction: identifying the main topics in a document for summarisation and tagging workflows.
• Conversational Language Understanding (CLU): the successor to LUIS. Intent classification and entity extraction for natural language interfaces — chatbots, command parsing, virtual assistants. CLU is trained on labelled utterances; the exam tests designing intents, entities, and training data strategies.
• Custom Question Answering: the successor to QnA Maker. Knowledge base built from FAQ documents, URLs, and manual QA pairs; serves grounded answers to natural language queries. Distinct from Azure OpenAI RAG: Custom QnA is deterministic and citation-based; Azure OpenAI RAG generates synthesised answers. Choose Custom QnA when auditability and exact-match retrieval are required; choose RAG when complex multi-document synthesis is needed.
• Azure AI Translator: neural machine translation (70+ languages), custom translation (domain-specific terminology via Custom Translator), document translation (preserving formatting), and real-time speech-to-translated-speech workflows in combination with Azure AI Speech.
• Azure AI Speech: speech-to-text (standard and custom acoustic models), text-to-speech (neural voices and custom neural voice), speaker recognition (verification and identification), and speech translation. Custom neural voice requires Limited Access approval for creating custom voices that resemble human speakers.

Domain 5 — Implement Knowledge Mining and Document Intelligence Solutions (~10–15%)

Azure AI Search: enterprise full-text and vector search over heterogeneous data sources. The AI enrichment pipeline is the exam-critical pattern: data source (Azure Blob Storage, SQL, Cosmos DB) → indexer → skillset (built-in cognitive skills + custom skills) → index (searchable fields with scoring profiles) → search interface. Know the built-in skills: OCR, language detection, entity recognition, key phrase extraction, image analysis, and sentiment. Custom skills are Azure Functions or REST endpoints that extend the pipeline with proprietary processing steps.

Vector search (added 2023): Azure AI Search supports storing and querying vector embeddings alongside traditional text fields. The exam tests hybrid search architectures that combine keyword BM25 ranking with semantic re-ranking using Azure OpenAI embeddings for RAG-based solutions.

Azure AI Document Intelligence: structured data extraction from documents. Pre-built models: invoice, receipt, business card, identity document, tax form (W-2, 1099), and general document. Custom models: template model (fixed-layout forms trained on 5–500 labelled samples) and neural model (variable-layout documents, more flexible, requires 50+ samples). The composed model endpoint aggregates multiple custom models and routes each document to the best-fit model automatically.

• The exam tests when Document Intelligence replaces OCR + custom parsing: when you need structured JSON output with labelled field names rather than raw text, and especially when the document structure is semi-structured (invoices from multiple vendors with different layouts).
• Knowledge store: Azure AI Search output projected into Azure Blob Storage or Azure Table Storage for downstream Power BI analysis or further ML processing. Tested when the requirement is enriched data persistence, not just search.

Domain 6 — Implement Generative AI Solutions (~15–20%)

The newest and highest-growth domain. Azure OpenAI Service gives enterprise access to OpenAI models (GPT-4o, GPT-4, GPT-3.5, Embeddings, DALL·E) under Microsoft’s data privacy terms — no training data use, private network endpoints, content filtering. The exam tests the Azure OpenAI programming model:

• System prompts: the instruction context that shapes model behaviour across all turns of a conversation. System prompt design is the primary tuning mechanism before fine-tuning is considered.
• Prompt engineering patterns: zero-shot (single instruction, no examples), few-shot (examples embedded in the prompt to guide output format), chain-of-thought (instructing step-by-step reasoning for complex problems), and self-consistency (sampling multiple completions and selecting the most common answer).
• Retrieval-Augmented Generation (RAG): the architecture for grounding LLM responses in proprietary data. Pattern: user query → embedding generation (Azure OpenAI Embeddings) → vector search (Azure AI Search) → top-k documents retrieved → injected into GPT prompt as context → grounded answer generated. RAG prevents hallucination on domain-specific questions and keeps training data private from the model.
• Semantic Kernel and Azure AI Foundry: Semantic Kernel is the Microsoft-backed SDK for orchestrating LLM calls, plugins, and memory in C#, Python, and Java applications. Azure AI Foundry (formerly Azure ML studio for AI) is the managed development hub for building, evaluating, and deploying generative AI applications.
• Content filtering: Azure OpenAI built-in safety filters (hate, violence, sexual, self-harm categories at severity thresholds). Custom blocklists for domain-specific term filtering. Groundedness evaluation: assessing whether model outputs are supported by the retrieved context.
• Fine-tuning vs RAG: the exam tests when each approach is appropriate. RAG for proprietary data retrieval, up-to-date information, and auditability. Fine-tuning for changing model style/tone, teaching domain-specific terminology when RAG context would be too large, and optimising for structured output formats.