AI-901 Study Guide

Fairness means an AI system treats everyone equitably and does not advantage or disadvantage groups of people. It matters most where AI informs decisions about people, such as loan approvals, hiring, or medical triage.

What you need to know

• Fairness is about avoiding bias, so the model does not favour or harm any group based on attributes like gender, ethnicity or age
• Bias usually enters through the training data, so representative data and testing across groups are the main defences
• Azure Machine Learning and the Responsible AI dashboard include fairness tools that compare a model's performance across groups
• A model that is accurate overall can still be unfair to a subgroup, which is why you measure per group

Microsoft Learn resource: Responsible AI (opens in a new tab)

Exam tip: If a scenario describes a model performing worse for one group of people than another, the principle is fairness.

Reliability and safety mean an AI system works consistently, as intended, and minimizes harm when conditions are unexpected. This is the principle behind rigorous testing before release, especially for systems that affect physical safety.

What you need to know

• AI systems must perform reliably even with unexpected or edge-case input
• Rigorous testing, monitoring, and clear operating limits are how you build reliability
• Safety matters most in high-stakes scenarios such as autonomous vehicles or health diagnostics
• Ongoing monitoring after deployment catches drift, where accuracy degrades as the real world changes

Microsoft Learn resource: Responsible AI (opens in a new tab)

Exam tip: Reliability and safety is the principle for "works consistently and predictably", especially where a failure could hurt someone.

Privacy and security mean an AI system respects the privacy of the people whose data it uses and is protected against misuse. AI depends on data, and much of that data is personal, so this principle governs how it is collected, stored, and accessed.

What you need to know

• Personal data must be protected both while training the model and while the deployed system runs
• People should have transparency and control over how their data is used
• Security controls such as encryption, access control, and network isolation protect the data and the model itself
• The principle covers new data the system generates about people, not only the data it was trained on

Microsoft Learn resource: Responsible AI (opens in a new tab)

Inclusiveness means AI should empower everyone and engage people across abilities, languages, and backgrounds. It pushes design to work for people who are often left out, including people with disabilities.

What you need to know

• AI should provide benefits regardless of physical ability, gender, ethnicity, or other factors
• Accessibility is a core part of inclusiveness, so features like speech, captioning, and alternative input matter
• Inclusive design considers a diverse range of people from the start rather than as an afterthought
• Supporting multiple languages and cultural contexts is part of reaching everyone

Microsoft Learn resource: Responsible AI (opens in a new tab)

Transparency means people understand how an AI system works, what it can and cannot do, and its limitations. Someone who understands a system's behaviour can use it appropriately and trust it correctly.

What you need to know

• People should be told when they are interacting with AI and what the system is for
• The system's capabilities and, just as important, its limitations should be made clear
• People should be able to understand the main factors behind a decision the AI influences
• Transparency includes being open about the data and the purpose behind the system

Microsoft Learn resource: Responsible AI (opens in a new tab)

Accountability means people, not the AI, are responsible for how an AI system behaves. The organizations that build and operate AI must answer for it and put governance around it.

What you need to know

• The people who design and deploy an AI system are accountable for how it operates
• Accountability is met through governance frameworks, organizational principles, and clear ownership
• It ensures AI meets ethical and legal standards, and that there is recourse when something goes wrong
• Microsoft's own Responsible AI Standard is an example of an accountability framework

Microsoft Learn resource: Responsible AI (opens in a new tab)

Exam tip: Accountability is about the humans and the governance around the system; if a question asks who is responsible for an AI's decisions, it is never the AI itself.

Generative AI models create new content, such as text, code, or images, in response to a natural language prompt. The text models behind them are large language models trained on huge amounts of text to predict the next token in a sequence.

What you need to know

• A token is the unit a model reads and generates, roughly a word or a piece of a word; prompts and responses are both measured in tokens
• Most models use the transformer architecture, which uses attention to weigh how words relate to each other
• The model does not look up answers; it predicts the most likely next token, which is why it can be fluent and still wrong
• Small language models (SLMs) are lighter and cheaper to run than large language models, trading some capability for speed and cost
• Multimodal models accept and produce more than one kind of content, such as text plus images

Microsoft Learn resource: Introduction to generative AI and agents (opens in a new tab)

Exam tip: When a question mentions predicting the next token or the transformer architecture, it is describing how a generative language model works.

Choosing a model means matching the job to a model's strengths: size, modality, cost, and whether it needs to reason or just respond. The Microsoft Foundry model catalog lists models from OpenAI, Microsoft, and the open-source community so you can compare them.

What you need to know

• The model catalog in Foundry is where you browse and compare models by capability, cost, and provider
• Bigger is not always better: a small model can beat a large one on a narrow task at a fraction of the cost
• Match the modality to what the app takes in and gives back, text, image, audio, or a mix
• Consider cost and latency, not just quality, because you pay per token and users wait for responses

Microsoft Learn resource: Get started with generative AI and agents in Azure (opens in a new tab)

Once you pick a model you deploy it to an endpoint and tune how it responds. The main parameters control randomness and length, and they turn up constantly in the exam's implementation questions.

What you need to know

• You deploy a model to get an endpoint and keys that your application calls
• Temperature (and top_p) control randomness: low values give focused, repeatable answers; high values give more varied, creative ones
• Max tokens caps the length of the response, and you pay per token for both input and output
• A system message sets the model's role and rules and applies to every request
• Deployment options range from standard, pay-per-token to provisioned throughput for steady, high volume

Microsoft Learn resource: Get started with generative AI and agents in Azure (opens in a new tab)

Exam tip: Low temperature for consistent, factual answers; high temperature for creative variety. That trade-off is a common question.

Microsoft groups Azure AI into a set of workloads, and the first skill is simply telling them apart from a described scenario. Each workload solves a different kind of problem.

What you need to know

• Generative AI creates new content, such as text, code, or images, from a prompt; agentic AI goes further, letting a model plan and take actions using tools
• Text analysis, part of natural language processing, reads text to find meaning: sentiment, key phrases, entities, and language
• Speech converts speech to text and text to speech, and translates spoken language
• Computer vision interprets images and video, and image-generation models create new images
• Information extraction pulls structured data out of documents, forms, images, audio, and video

Microsoft Learn resource: Introduction to AI concepts (opens in a new tab)

Exam tip: Read the scenario for the input and the output. "Turn this recording into text" is speech; "find the total on this invoice" is information extraction.

Text analysis, part of natural language processing, turns unstructured text into structured insight. Azure Language in Foundry Tools provides these techniques as ready-to-call features.

What you need to know

• Key phrase extraction pulls the main points out of a block of text
• Entity detection, also called named entity recognition, finds and categorizes things like people, places, dates, and organizations
• Sentiment analysis scores text as positive, negative, or neutral, with a confidence score
• Summarization produces a short version of a longer document, either by extracting key sentences or generating new ones
• Language detection identifies which language a piece of text is written in

Microsoft Learn resource: Introduction to natural language processing concepts (opens in a new tab)

Exam tip: Match the verb: "categorize the people and places" is entity detection; "is this review happy or angry" is sentiment analysis.

Speech covers two opposite directions: turning spoken audio into text, and turning text into spoken audio. Azure Speech in Foundry Tools provides both, plus translation of speech.

What you need to know

• Speech to text can transcribe in real time or from recorded audio files
• Text to speech offers a range of natural, configurable neural voices
• Azure Speech also does speech translation, recognizing one language and producing another
• A multimodal model can now handle speech directly, which is the newer path the exam expects you to know

Microsoft Learn resource: Speech (opens in a new tab)

Exam tip: Speech to text is recognition (audio in, text out); text to speech is synthesis (text in, audio out). Do not mix the directions.

Computer vision interprets the content of images and video; image-generation models do the reverse, creating images from a text prompt. Both are part of the vision workload.

What you need to know

• Image classification assigns a label to a whole image; object detection finds and locates multiple objects with bounding boxes
• Optical character recognition (OCR) reads printed and handwritten text from images
• Face detection locates faces and can support verification scenarios, under Microsoft's responsible AI limits
• Image analysis generates captions, tags, and detected objects for an image
• Image-generation models such as DALL-E create new images from a natural language description

Microsoft Learn resource: Introduction to computer vision concepts (opens in a new tab)

Exam tip: Classification labels the whole image; object detection also tells you where each object is.

Information extraction pulls structured data out of unstructured content, whatever the format. It is how you turn a pile of invoices, recordings, or scanned forms into fields you can use.

What you need to know

• From documents and forms, extract fields, key-value pairs, and tables, such as an invoice's total or a form's answers
• From images, extract text with OCR along with identified objects or content
• From audio and video, extract a transcript first, then insights like topics, sentiment, and named people
• Azure Content Understanding is the Foundry service that handles extraction across all these content types in one place
• The output is structured data, fields or JSON, that an application or workflow can act on

Microsoft Learn resource: Introduction to AI-powered information extraction concepts (opens in a new tab)

Exam tip: Information extraction is about getting structured fields out; if the scenario ends with "into a spreadsheet" or "as fields", that is the workload.

A prompt is how you steer a generative model, and the exam expects you to know the difference between the system prompt and the user prompt. Good prompting is often the difference between a useful answer and a useless one.

What you need to know

• The system prompt, or system message, sets the model's role, tone, and rules, and applies to the whole conversation
• The user prompt is the specific request or question for a single turn
• Clear instructions, context, and examples in the prompt improve the response; this is prompt engineering
• One-shot and few-shot prompting give the model one or a few examples to follow
• Grounding the prompt with your own data reduces made-up answers

Microsoft Learn resource: Get started with generative AI and agents in Azure (opens in a new tab)

Exam tip: The system message shapes every response; the user message is the individual turn. Questions test that you know which is which.

Before you can call a model you deploy it from the Foundry model catalog, then test it in the portal's playground. This is usually the first hands-on step in any Foundry project.

What you need to know

• You choose a model in the model catalog and deploy it to get an endpoint and keys
• The playground lets you test prompts and settings without writing any code
• In the playground you can set the system message and adjust parameters like temperature
• A Foundry project groups your deployments, data, and other resources together
• The endpoint and key from the deployment are what your application code will use next

Microsoft Learn resource: Get started with generative AI and agents in Azure (opens in a new tab)

Exam tip: The playground is for testing in the portal; the endpoint and key are for calling the model from code.

Once a model is deployed, a few lines of code turn it into a chat app. The exam shows short snippets and asks what they do or what is missing.

What you need to know

• You install the Foundry SDK and authenticate with the project endpoint and a key or a Microsoft Entra credential
• You create a client, send a list of messages (system, then user), and read the model's reply
• The SDK is available for Python and other languages; AI-901 assumes Python syntax
• Conversation history is sent with each request, because the model itself is stateless between calls
• Keys and endpoints belong in configuration or environment variables, never hard-coded in the app

Microsoft Learn resource: Plan and prepare to develop AI solutions on Azure (opens in a new tab)

Exam tip: The model does not remember previous turns on its own; your app resends the history on each call.

An agent is a model given instructions, tools, and sometimes knowledge, so it can carry out a task rather than just answer. Foundry lets you build and test one in the portal.

What you need to know

• An agent combines a model, a set of instructions, and tools it is allowed to use
• You give the agent instructions, its purpose and rules, much like a system prompt
• Tools let an agent do things such as run code, call an API, or search your data
• Knowledge, or grounding data, lets the agent answer from your content, not just its training
• You can test the agent in the portal before connecting a client application to it

Microsoft Learn resource: Get started with generative AI and agents in Azure (opens in a new tab)

Exam tip: An agent is a model plus instructions plus tools; a plain chat completion is just the model answering.

After building an agent in the portal, you connect an application to it with the SDK, much like a chat client but pointed at the agent.

What you need to know

• The app connects to the agent through the Foundry SDK using the project endpoint and credentials
• You start or reuse a conversation thread, send the user's message, and read the agent's response
• The agent, not your code, decides when to call its tools to complete the request
• Authentication uses a key or, better, Microsoft Entra ID with role-based access
• The same responsible AI and content safety considerations apply to an agent app

Microsoft Learn resource: Plan and prepare to develop AI solutions on Azure (opens in a new tab)

Text analysis features from Azure Language in Foundry Tools plug into an app with the SDK, so you can score sentiment or pull entities out of user text.

What you need to know

• Azure Language provides sentiment analysis, key phrase extraction, entity recognition, language detection, and summarization
• You call the service through its endpoint and key, or through the Foundry SDK
• Each feature returns structured results, often with a confidence score you can act on
• Sending several documents in one call is more efficient than one call each
• No model training is needed for these prebuilt features; you call them directly

Microsoft Learn resource: Get started with text analysis in Azure (opens in a new tab)

Exam tip: These Language features are prebuilt, so reach for training only when a custom entity or classification is genuinely needed.

A multimodal model can take audio in and respond, so an app can answer a spoken question directly. This is the newer, model-based path for speech scenarios.

What you need to know

• A multimodal model accepts more than text, including audio, in the same request
• The app captures the user's audio and sends it to the deployed model
• The model can respond with text, or with audio for a spoken reply
• This path can replace chaining separate speech-to-text, language, and text-to-speech steps
• You still deploy the model in Foundry and call it through an endpoint

Microsoft Learn resource: Get started with generative AI and agents in Azure (opens in a new tab)

For dedicated speech work, Azure Speech in Foundry Tools gives you transcription and synthesis through the SDK.

What you need to know

• Speech to text transcribes microphone or file audio into text
• Text to speech turns text into natural, neural-voice audio
• The Speech SDK handles real-time streaming as well as one-shot calls
• You can choose the voice, the language, and the output format for synthesis
• Speech translation recognizes one language and outputs another

Microsoft Learn resource: Get started with speech in Azure (opens in a new tab)

Exam tip: For a dedicated transcription or voice feature, use Azure Speech; for a conversational audio assistant, a multimodal model may be simpler.

A multimodal model can accept an image alongside text, so your app can ask questions about a picture.

What you need to know

• You send an image plus a text prompt in the same request to a multimodal model
• The model can describe the image, answer questions about it, or read text and details from it
• The image can be passed as a URL or as base64-encoded data
• This replaces training a custom vision model for many one-off "what is in this image" tasks
• Clear prompts, such as "list every product on this shelf", get better structured answers

Microsoft Learn resource: Get started with generative AI and agents in Azure (opens in a new tab)

Image-generation models create original images from a text description, and Foundry lets you deploy and call them like any other model.

What you need to know

• An image-generation model such as DALL-E produces an image from a natural language prompt
• You describe the subject, style, and composition in the prompt; more detail gives more control
• The model returns an image, or a URL to one, that your app can display or store
• Generated images are subject to content safety filters and responsible AI limits
• You deploy the image model in Foundry and call it through an endpoint, like a text model

Microsoft Learn resource: Get started with generative AI and agents in Azure (opens in a new tab)

Exam tip: Computer vision reads existing images; image-generation models create new ones. Watch which direction the scenario asks for.

Azure Vision in Foundry Tools adds prebuilt image features, captioning, OCR, and object detection, to an app through the SDK.

What you need to know

• Image analysis returns captions, tags, and detected objects for an image
• OCR, through the Read capability, extracts printed and handwritten text from images and documents
• Object detection returns objects with bounding-box coordinates
• The service is prebuilt, so no training is needed for common tasks
• Custom models are available when the prebuilt categories are not enough

Microsoft Learn resource: Get started with computer vision in Azure (opens in a new tab)

Azure Content Understanding pulls structured fields out of documents and forms, so invoices, receipts, and contracts become data. It is the Foundry service the exam names for this.

What you need to know

• Azure Content Understanding extracts fields, key-value pairs, and tables from documents
• Prebuilt analyzers handle common document types such as invoices and receipts out of the box
• Custom analyzers let you define the fields to pull from your own form layouts
• The output is structured, fields and JSON, ready for an application or workflow
• One service handles multiple content types, documents, images, audio, and video

Microsoft Learn resource: Get started with AI-powered information extraction in Azure (opens in a new tab)

Exam tip: Content Understanding is the name to remember for pulling fields out of documents in Foundry.

Content Understanding also reads images, extracting the text and details you define.

What you need to know

• It reads printed and handwritten text from images using OCR
• It can return described content and specified fields from a photo or a scan
• You point it at an image and define the schema of what to pull out
• Results come back as structured data, often with confidence scores
• It is the same service used for documents, applied to image input

Microsoft Learn resource: Get started with AI-powered information extraction in Azure (opens in a new tab)

For recordings, Content Understanding produces a transcript and then the insights you ask for.

What you need to know

• It transcribes audio and video into text as the first step
• From the transcript it can extract topics, key points, and named entities
• It can identify speakers and segment a recording
• The extracted insights come back as structured fields, not just raw text
• This turns meetings, calls, and clips into searchable, structured data

Microsoft Learn resource: Get started with AI-powered information extraction in Azure (opens in a new tab)

Wiring Content Understanding into an app is the same SDK pattern: send content, get structured results.

What you need to know

• The app sends the document, image, audio, or video to the Content Understanding endpoint
• You choose a prebuilt analyzer or your own custom one for the content type
• The service returns structured fields the app can store or route
• Authenticate with a key or Microsoft Entra ID, and keep secrets out of code
• Large files are processed asynchronously, so the app waits for or polls for the result

Microsoft Learn resource: Get started with AI-powered information extraction in Azure (opens in a new tab)

Exam tip: The pattern is always send content, pick an analyzer, read structured fields back.