AI-901 Concept Map (Mind Map): Section 1

Every topic from section 1 of the exam on one page, with the pairs the exam swaps to trick you. Section 1 counts for 40-45% of your score.

1.1 Principles of responsible AI

Six principles, and the controls that put them into practice

The six principles

Fairness: AI should treat similar people the same way: no group gets worse outcomes or more errors because of who they are. You check it by comparing results group by group.
Reliability and safety: AI should work as intended, cope with unusual or bad input, and fail safely, with a person stepping in when an error could cause harm.
Privacy and security: Personal data is collected only when needed, protected, kept only as long as needed and seen only by people allowed to see it. Access to the system itself is secured too.
Inclusiveness: AI should empower everyone and shut nobody out, including people with disabilities, strong accents or other languages. It often means offering more than one way in, such as text as well as voice.
Transparency: People know they are dealing with AI, what it does, what data it uses and what its limits are, and its decisions can be understood.
Accountability: People and organizations stay responsible for their AI: named owners, governance, review before launch and a way to switch a system off. The company that uses a model for decisions answers for them, even with a third-party model.

Putting principles into practice

Per-group evaluation: Measure accuracy, approvals and error rates separately for each group of people on a test set, not just overall. When a group is under-represented, add data and measure again.
Edge-case and adversarial testing: Test with ambiguous, off-topic, misleading and hostile input, not only the happy path, and check that the system fails safely. Red teaming is the organized version.
Human in the loop: A qualified person reviews or approves the AI output before it takes effect, for high-impact decisions or when errors look like right answers.
Confidence threshold: Act automatically only when the model’s confidence score is above a set level; otherwise stop and send the case to a person.
Data minimization and redaction: Send and keep only the personal data a task needs: detect and redact personal data before text reaches a model, and retrieve only the records each user may see.
AI disclosure: Say clearly, where people see it, that they are talking to an AI or looking at AI-generated content, and state its limits.
Guardrails (content filters): Safety checks on a Foundry model deployment that screen prompts and responses for harmful content such as hate, violence, sexual content and self-harm, and for prompt attacks. Deployed models start with default guardrails you can adjust.

1.2 AI model components and configurations

How generative models work, which to pick, how to deploy and tune them

How a generative model works

Token: The unit a model reads, limits and bills: a whole word, part of a word or a punctuation mark. A text usually has more tokens than words.
Embedding: A list of numbers (a vector) that captures meaning: texts with similar meaning get vectors that sit close together. It powers search by meaning and retrieval.
Attention (transformer): The mechanism inside a transformer, the network design behind language models: each token weighs the tokens around it, so bank takes its meaning from river or money nearby.
Next-token prediction: A language model writes by predicting the most likely next token, adding it and repeating. The text is fluent by design, but factual only when its training or its prompt supports it.
Hallucination: A plausible-sounding answer that is wrong or invented, often about facts the model never saw, such as your company’s own policies.
Context window: The maximum number of tokens one request can hold: instructions, conversation history, retrieved text and the answer together.
Stateless calls (conversation history): Each call to a model is independent. To hold a conversation, the app resends the earlier messages or links each call to the previous response.

Pick the model by capability

Large language model (LLM): A large, general-purpose text model for complex, open-ended language work, built on a foundation model pretrained on broad data.
Small language model (SLM): A compact model, such as Microsoft’s Phi family, for focused tasks with low latency and cost, even offline on the device itself.
Reasoning model: A model that works through a problem step by step before answering: best for math, logic, planning and multi-step problems, but slower and costlier per answer.
Multimodal model: Takes several input types together, such as text with images or audio, and reasons over them. One that reads images is also called vision-enabled.
Embedding model: Turns text into embeddings for search and retrieval, for example text-embedding-3-small. It does not chat or write answers.
Image generation model: Creates new images from a text prompt, and some also edit existing images, for example the gpt-image series in Foundry.
Model catalog: The Foundry Models catalog: find, filter, compare and deploy models, with model cards, benchmarks and leaderboards.
Models sold by Azure: Models Microsoft hosts and sells under Microsoft Product Terms, with Microsoft support and SLAs. They include the Azure OpenAI models.
Models from partners and community: Most of the catalog: models from third-party providers, research labs and the community, offered under each provider’s terms.

Ground or customize

Retrieval-augmented generation (RAG): The app finds the relevant passages of your documents (split into chunks and indexed) and adds them to the prompt, so the model answers from current, private facts.
Fine-tuning: Further training a model on your own examples to change its behavior, style or output format for a task.

Deploy it

Global Standard: Pay per token; requests may be processed in any Azure region. The default choice: newest models first, lowest price, highest quota.
Data Zone Standard: Pay per token; processing stays inside a Microsoft data zone (US, EU or Asia Pacific), with more quota than the single-geography Standard type.
Provisioned throughput: Reserved capacity, bought in provisioned throughput units (PTU), for steady, high-volume, latency-sensitive traffic. Comes as Global, Data Zone and Regional Provisioned.
Global Batch: Large jobs sent as a file and processed asynchronously, with a 24-hour target turnaround, at a 50% discount on the Standard price.
Tokens-per-minute (TPM) quota: The rate limit given to a pay-per-token deployment from your subscription’s quota. Going over it returns rate-limit (429) errors.

Tune the output

Temperature: Controls randomness: low gives consistent, focused answers; higher gives more varied, creative ones.
Max output tokens: The longest answer the model may return. Set too low, replies stop mid-sentence.
Reasoning effort: On reasoning models, sets how much the model thinks before answering, such as low, medium or high: less is faster and cheaper, more handles harder problems.

1.3 AI workloads

Generative and agentic AI, text, speech, vision and information extraction

Generative and agentic AI

Generative AI: AI that creates new content, such as text, code, images or audio, from a prompt.
AI agent: A model plus instructions plus tools, so it can carry out multi-step tasks and take actions, not just reply.
Knowledge tool: A tool that lets an agent read information, such as a policy document or a search index, to ground its answers.
Action tool: A tool that lets an agent do something in another system, such as booking leave, creating a shipment or sending an email.
Multi-agent system: Several specialist agents, for example research, drafting and review, that work together instead of one agent doing every job.
Indirect prompt injection: Instructions hidden in a document, email or web page that an agent reads, trying to make it act. Retrieved text must be treated as data, never as orders.

Text analysis

Keyword extraction: Returns the main terms and topics of a text, with no type attached. Azure Language calls it key phrase extraction.
Entity detection (NER): Named entity recognition finds named items in text and labels their type: person, place, organization, date, quantity.
Sentiment analysis: Scores text as positive, negative, neutral or mixed, for the whole document and per sentence.
Extractive summarization: Builds a summary by picking the most representative sentences and keeping them unchanged.
Abstractive summarization: Writes new, shorter sentences that capture the main points, the way a person would.
Language detection: Identifies the language of a text, with a confidence score, so a pipeline can route each text before other analysis.
PII detection: Finds personal data such as names, phone numbers and card numbers, and can return a redacted copy of the text.
Lemmatization: Reduces each word to its dictionary form using vocabulary and grammar, so better becomes good and ran becomes run.

Speech

Speech recognition (speech to text): Turns spoken audio into text.
Speech synthesis (text to speech): Turns text into natural-sounding speech with a neural voice.
Real-time transcription: Transcribes live audio as it is spoken, for live captions and voice commands.
Batch transcription: Transcribes large volumes of stored recordings asynchronously.
Speech translation: Translates spoken audio into text or speech in other languages, in real time.
Custom speech: Adapts the recognition model with your own text and audio, for specialist vocabulary, accents or noisy places.
Prosody: The rhythm, stress and intonation of synthesized speech: the melody of the voice.
Text normalization: The synthesis step that turns symbols, numbers, dates and abbreviations into words to say, such as $5 into five dollars.
Voice Live: Azure Speech feature for real-time voice conversations with a generative model or agent: recognition, model and synthesis in one low-latency managed service, with natural turn taking.

Computer vision and image generation

Image classification: Gives one label to a whole image, such as this photo shows a damaged part.
Object detection: Finds each object in an image and returns its label and a bounding box.
Semantic segmentation: Labels each pixel of an image with the class it belongs to, giving exact shapes.
Optical character recognition (OCR): Reads printed or handwritten text in images and documents, with the position of each piece.

Information extraction

Field extraction: Turns content into named values defined in a schema (invoice number, total, date), matched by meaning rather than by the label printed on the page.
Template-based extraction: Reads fields at known positions on one fixed form layout: simple and easy to explain, but it breaks when the layout changes.
Content Understanding: Azure Content Understanding in Foundry Tools: uses generative AI to turn documents, images, audio and video into structured output, such as JSON fields and, for documents, confidence scores and grounding.

Part of the Microsoft Azure AI Fundamentals (AI-901) Study Kit: Concept Maps, Flashcards and practice questions. Also free: the Section 1 flashcards. New to the exam? Read the AI-901 study guide.