AI-901 · CHEAT SHEET

Microsoft Certified: Azure AI Fundamentals

Key concepts grouped by official exam weight.

Download as PDF (free, no sign-up)

1. Identify AI concepts and capabilities

40–45%

Describe principles of responsible AI

Fairness means an AI system must not treat similar people or groups differently. Because people choose the data and it can carry unconscious bias, developers minimize bias in training data and test results across sensitive groups, for example with the fairness assessment in the Responsible AI dashboard.
AI is based on probabilistic models and is not infallible, so applications must mitigate its risks: work as designed, respond safely to unexpected conditions and avoid acting when confidence is low. Content filters help, but the principles apply from design through operation.
Models are trained on data that may include personal information, so that data must be kept secure and the trained model must not reveal private details. This is applied with access control, encryption in transit and at rest, auditing, deleting data no longer needed, and tools such as SmartNoise (differential privacy) and Counterfit.
Inclusiveness means the potential of AI is open to everyone and solutions do not exclude any user. In practice, a speech-based agent also offers text captions, and agent design considers accessibility, language variation and different levels of technical proficiency, with input from diverse communities from the start.
Transparency means users know how an AI system works and what limitations it has: stating that AI is used, describing features of the training data without revealing confidential information, and providing Transparency Notes or in-product disclosures. Interpretability (global, local and cohort explanations) helps people understand its decisions.

Identify AI model components and configurations

Generative language models split text into tokens, turn them into embeddings with a transformer (an encoder using attention), and a decoder predicts the most probable continuation of a prompt one token at a time. Examples in a prompt condition only that inference, and the app resends history to give context.
Choosing a model starts with the task: embeddings for semantic search, multimodal models to analyze images with text, image or video models to generate content, small models for lightweight apps and fast models for high-volume chat. For a well-defined case, a Foundry tool with prebuilt models gives more predictable results.
When you deploy a model you choose the deployment type (Global Standard by default, Data Zone or Standard for residency, Provisioned for reserved capacity, Batch for asynchronous jobs), the version and the tokens per minute. Temperature, top_p and max output tokens are adjusted in the playground or in code.

Identify AI workloads

Common workloads are generative AI and agents (a model, instructions and tools that act), text analysis (sentiment, key phrases, entities, PII), speech (recognition, synthesis and speech-to-speech), vision (image analysis and generation), and information extraction from documents, audio and video using schemas.
Key phrase extraction lists the main concepts of a text, entity recognition detects and classifies items such as people, places, dates or quantities, and sentiment analysis labels sentences and documents as positive, neutral or negative with confidence scores. Summarization can be extractive (original sentences) or abstractive (newly written wording).
Speech recognition (speech to text) converts audio into text through an acoustic model and a language model, and can run in real time, fast or in batch, with custom models for specific vocabulary. Speech synthesis (text to speech) produces audio with neural voices and is tuned with SSML. Voice Live joins both for real-time conversations with an agent.
Vision-enabled multimodal models accept images and text in one request (as URL or base64) to describe photos and answer questions about them. Image generation models create pictures from text, and video generation models produce clips through asynchronous jobs. Azure Vision adds OCR and Face for reading text and analyzing faces.
Azure Content Understanding in Foundry Tools turns documents, images, audio and video into structured data through analyzers with a field schema, going beyond basic OCR. Audio analyzers transcribe with diarization, and video analyzers add key frames, shots and segments. Plain text is analyzed with Azure Language, for example with entity recognition.

2. Implement AI solutions by using Microsoft Foundry

55–60%

Implement generative AI apps and agents by using Foundry

The system prompt sets the assistant role, tone, scope and guardrails, while the user prompt is the specific request. A good prompt is clear, specific and concise, states the output format, can include examples (few-shot), grounding data and an out when there is no answer, and repeats key instructions at the end when the context is long.
Deploying a model in the Foundry portal requires a project and a deployment type choice; the deployment name is the value you pass in the model parameter. When it finishes you land in the playground, where you test prompts, adjust parameters such as temperature, compare up to three models and check the Code tab.
A lightweight chat client collects user input, calls the deployed model and displays the response. With the Foundry SDK for Python you create an AIProjectClient with the project endpoint and DefaultAzureCredential, get the OpenAI client with get_openai_client and call responses.create with model and input.
A Foundry agent combines a model, instructions and tools; knowledge adds context through RAG. In the agents playground you configure and test conversations without code, save changes as immutable versions, and use tracing, evaluations and publishing to debug, measure and share the agent.
To call an agent from a lightweight Python client, you create an AIProjectClient with the project endpoint and get an OpenAI client with get_openai_client, pointing at the agent by name. A conversation plus responses.create keeps history across turns; the recommended pattern is to create the agent once and reuse it.

Implement AI solutions for text and speech by using Foundry

A lightweight app can analyze text with a general-purpose model (flexible, probabilistic output) or with Azure Language in Foundry Tools (analyzers with structured, deterministic output). With the Python SDK you use TextAnalyticsClient and methods such as detect_language, recognize_pii_entities, analyze_sentiment or extract_key_phrases.
An audio-enabled chat model deployed in Foundry receives the spoken question as an input_audio item holding base64 data and its format. To get a spoken reply you set the text and audio modalities plus a voice and format; the returned audio is base64 and must be decoded. The separate realtime API covers low-latency interactive voice.
With the Azure Speech SDK for Python you create a SpeechConfig with the Foundry resource endpoint and key, a SpeechRecognizer for speech-to-text and a SpeechSynthesizer for text-to-speech. Transcription can be real-time or batch (asynchronous), and Voice Live combines recognition, a generative model and synthesis for voice agents.

Implement AI solutions with computer vision and image-generation capabilities by using Foundry

A vision-enabled multimodal model deployed in Foundry accepts text and images in one request and can describe photos, answer questions about them or interpret charts and screenshots. Images are sent as a publicly accessible URL or as base64 data; the detail parameter tunes the level of detail and max_tokens prevents truncated replies.
Foundry image generation models create images from text and can also edit or transform existing images; they are deployed in a Foundry resource, tried in the playground and used through the Responses API or the images API, which returns base64 data. Video generation is an asynchronous job: create, poll the status and download.
A lightweight vision app uses the OpenAI Python SDK with the Foundry resource key and endpoint (or Entra ID) and passes the deployment name as model. It builds a multi-part message with text and image and reads the answer from output_text; everything can first be tested in the playground, which offers sample code.

Implement AI solutions for information extraction by using Foundry

Content Understanding processes documents with an analyzer that combines content extraction (OCR, tables, sections, selection marks, barcodes) and schema-based field extraction, applied semantically. Fields use the extract, classify or generate methods, and confidence and grounding are enabled with estimateFieldSourceAndConfidence. contentCategories with enableSegment classifies and splits mixed files.
Content Understanding image analyzers turn photos into structured data from a schema of fields, descriptions and output types, with uses such as quality control or shelf analysis. Their fields use generate or classify, because extract only exists for documents. If the goal is mainly text, a document schema fits better; prebuilt-imageSearch generates a description for search.
Audio analyzers transcribe in WebVTT with diarization, role detection, language detection and multilingual transcription, and extract fields such as summary or sentiment; prebuilt-callCenter and prebuilt-audioSearch are the prebuilt options. For video, content extraction provides transcript, shots and key frames, and a generative model adds fields and segmentation through enableSegment; prebuilt-videoSearch is the RAG analyzer.
A lightweight app uses the azure-ai-contentunderstanding SDK: it creates a ContentUnderstandingClient with the Foundry resource endpoint and a credential, and calls begin_analyze with an analyzer and its inputs. Analysis is asynchronous: the poller waits for the result, with markdown and fields with confidence; in REST you poll Operation-Location. Custom analyzers are created with PUT.

Practice this certification