Azure AI Engineer Associate (AI-102) Cheat Sheet

by

What You’re Actually Being Tested On

AI-102 is often misunderstood.
It’s not testing whether you can design advanced AI systems. It’s testing whether you can choose the correct Azure AI service and wire it up properly. Most wrong answers happen because people over complicate things. This post is a straight-to-the-point cheat sheet based on how the exam questions are written and scored.

One Rule That Solves Most Questions

Pick the most Azure-native, least complex, and most secure solution that meets the requirement. When reviewing answer options, eliminate anything that introduces custom training without a clear need, hard-codes secrets or keys, uses the wrong data type (text vs image vs audio), or feels overengineered. If one option feels boring and obvious, that’s usually the correct one.

You Must Know Python or C#

This is not optional.

A large part of the exam assumes you understand how to call Azure AI services from code. This includes creating clients, authenticating (preferably with Managed Identity), calling SDK methods, and handling inputs and outputs correctly.

You don’t need to memorize full programs, but you must be comfortable reading and reasoning about Python SDK examples or C# (.NET) SDK examples. If you don’t recognize how services like Vision, Language, OpenAI, or Document Intelligence are called from code, many questions will feel confusing even if you know the theory. If you’re short on time, pick one language (Python or C#) and stick to it.

Azure OpenAI Service

Use this only when the output is generative.

Typical use cases include chatbots and copilots, text generation and summarization, embeddings and semantic search, and code generation. Exam signals include LLMs, prompts, chatbots, Copilot-style scenarios, natural language responses, and few-shot or zero-shot examples. You should know the available models (GPT-4, GPT-4o, GPT-35-Turbo, embeddings), that content moderation exists and matters, how system, user, and assistant roles work, and that smaller models are cheaper and often sufficient. Managed Identity is the preferred authentication method.

Python client:

  • AzureOpenAI (from the openai package)

You should recognize patterns like:

  • Creating an AzureOpenAI client

  • Calling chat.completions.create()

  • Calling embeddings.create()

A common mistake is choosing Azure AI Language when the question clearly expects generated text or chat behavior.

Azure AI Language

Use this when you are analyzing text, not creating it.

This service is used for sentiment analysis, key phrase extraction, entity recognition (NER), PII detection, language detection, and question answering. Custom options include custom text classification and custom NER, both of which require labeled training data. Exam signals are phrases like analyze text, extract entities, detect PII, or classify documents.

Python client:

  • TextAnalyticsClient

Typical operations:

  • analyze_sentiment

  • recognize_entities

  • recognize_pii_entities

  • extract_key_phrases

If no generative output is required, Azure OpenAI is usually the wrong answer.

Azure AI Vision

This service is for images and OCR, not business documents.

It is used for OCR (printed and handwritten), image tagging, object detection, and face detection (not recognition). OCR uses the Read API. No training is required for standard scenarios, and it works with images and PDFs.

Python client:

  • ImageAnalysisClient

Key thing to recognize:

  • OCR is done via the Read visual feature

A common mistake is using Vision OCR when the question clearly describes invoices, receipts, or structured forms.

Azure AI Document Intelligence

Use this when structure matters.

It is designed for invoices, receipts, forms, tables, and key-value pairs. Exam signals include structured fields, business documents, and PDFs with layout. You should default to prebuilt models such as Invoice, Receipt, or ID. Custom models are only needed for unique layouts.

Python client:

  • DocumentAnalysisClient

Common exam signal:

  • model_id="prebuilt-invoice" (or receipt / ID)

For business documents, this service is preferred over Vision.

Azure AI Speech

Any time audio is involved, this service should come to mind. It is used for speech-to-text, text-to-speech, and speech translation. Exam signals include audio input, call center scenarios, transcription, or spoken output.

Python client:

  • SpeechRecognizer

  • SpeechSynthesizer

From the azure.cognitiveservices.speech package.

Custom speech is rarely the correct answer unless the question explicitly requires it.

Azure AI Translator

This service is simple and focused.

It is used for text translation, document translation, and language detection. No training is required, and both real-time and batch processing are supported.

Python usage pattern:

  • REST API via requests

  • No dedicated high-level SDK like Language or Vision

Recognize:

  • /translate endpoint

  • Language codes (to=, from=)

It is often combined with Speech or Language services.

Azure AI Content Safety

If the question mentions safety, moderation, or compliance, this service must be included.

It is used to detect harmful or abusive content in text and images.  Content Safety is built into Azure OpenAI but can also be used as a standalone service. Ignoring safety requirements is an easy way to lose points.

Python client:

  • ContentSafetyClient

Recognize:

  • Text or image analysis for harmful categories

  • Often paired with Azure OpenAI

Azure AI Search

This service is about retrieval, not generation.

It is used for document search, indexing structured and unstructured data, and vector search using embeddings. It is commonly used in RAG architectures together with Azure OpenAI.

Python client:

  • SearchClient

Common pattern:

  • client.search(...)

OpenAI does not replace search.

Azure Machine Learning

This is almost never the right answer.

Only choose Azure Machine Learning if the question explicitly states that you must train your own model, build a custom algorithm, or manage a full ML lifecycle.

If a managed AI service can solve the problem, Azure Machine Learning is probably wrong.

Fast Service Mapping

  1. Chat, prompts, and LLM scenarios map to Azure OpenAI.
  2. Sentiment analysis, entities, and PII map to Azure AI Language.
  3. Custom text classification or NER maps to Language (Custom).
  4. OCR and image analysis map to Azure AI Vision.
  5. Invoices, receipts, and forms map to Document Intelligence.
  6. Audio input or output maps to Azure AI Speech.
  7. Translation maps to Azure AI Translator.
  8. Moderation maps to Content Safety.
  9. Search and RAG map to Azure AI Search.
  10. Custom ML training maps to Azure Machine Learning.

Authentication and Security: Correct answers include Managed Identity, Azure Key Vault, and Entra ID.  Incorrect answers include secrets in code, plain API keys, and hard-coded credentials.

Common Error Clues: If you see 401 or 404 errors, request failures, or region-specific hints, the issue is often a region or endpoint mismatch.

SDK vs REST: SDK usage is the default answer. REST is only correct if the SDK does not support the language in question.

Cost Awareness Still Matters: Good answers include prebuilt models, batch processing, caching, and smaller OpenAI models. Bad answers include unnecessary custom training and always-on compute.

How to Approach Case Studies: Read the requirements first. Identify data type, latency, and security needs. Ignore filler text.

Final Advice: AI-102 is a design and configuration exam, not a data science exam. If the solution feels simple, it’s probably the one Microsoft expects.

N/B

This material is based on my own AI-102 exam prep notes, compiled from Microsoft documentation and hands-on study. Some sections were refined with the help of ChatGPT for structure and readability. This is not official Microsoft material.

Leave a Comment