What does AI-901 Domain 2 cover?

Domain 2 covers four implementation areas: generative AI apps and agents using the Foundry portal and SDK (prompts, model deployment, lightweight chat clients, single-agent solutions); text and speech solutions using Foundry (text analysis apps, multimodal speech responses, Azure Speech in Foundry Tools); computer vision and image generation using Foundry (visual input interpretation, image generation, vision-capable applications); and information extraction using Azure Content Understanding (extracting data from documents, images, audio, and video). It accounts for 55–60% of the AI-901 exam.

What is a single-agent solution in AI-901?

A single-agent solution in AI-901 refers to building an AI agent in the Microsoft Foundry portal that can carry out automated, multi-step tasks. The exam tests your ability to create and test a single-agent solution in the Foundry portal and build a lightweight client application that interacts with that agent. This reflects the growing importance of agentic AI patterns in real-world Azure AI development.

What is Azure Content Understanding in AI-901 Domain 2?

Azure Content Understanding is a Foundry-native service that extracts structured information from unstructured content including documents and forms, images, audio, and video. Domain 2 tests your ability to use Content Understanding for each media type and to build lightweight applications with information-extraction capabilities. It is a key differentiator of AI-901 compared to the older AI-900 syllabus.

AI-901 Domain 2: Implement AI Solutions Using Microsoft Foundry (55

📌

Implement AI Solutions by Using Microsoft Foundry domain is organised into four skill areas, each covering a distinct implementation context within Microsoft AI Foundry: generative AI apps and agents; text and speech solutions; computer vision and image generation; and information extraction using Azure Content Understanding in Foundry Tools. Questions mix scenario-based decisions, True/False statements about specific behaviours, sentence-completion questions, and Python code-completion questions, where candidates fill in a single value in a short code snippet. The table below maps every official sub-topic to the type of question you will face on it in the exam.

What the Exam Tests

Sub-topic / Area	What kind of question to expect
Create effective system and user prompts for generative AI models	True/False questions test what a system prompt does and does not do. Candidates need to understand the specific role a system prompt plays in a generative AI application and be able to distinguish it from other configuration mechanisms such as resource quotas, authentication controls, and deployment settings. Multi-select questions present a list of possible purposes for instructions when prompting a generative AI model and ask candidates to identify the correct ones. Candidates need to know what instructions actually govern versus what is controlled elsewhere in the Foundry configuration. Sentence-completion questions describe a requirement around defining an agent's role and behaviour and ask candidates to identify which component must be configured to achieve this.
Deploy a model and interact with it in the Foundry portal	Scenario questions describe a task within the Foundry portal and ask candidates to identify which area or feature of the portal is used to accomplish it. Candidates need to know the purpose of each named section of the portal and be able to distinguish between sections with similar-sounding names. True/False questions test what the Foundry portal requires and allows at different points in the model lifecycle, including what actions are and are not prerequisites for deployment and testing. True/False questions also test the capabilities and limitations of Evaluators in Foundry, including what Evaluators can assess and what they cannot do automatically.
Lightweight chat client application using the Foundry SDK	Code-completion questions present a short Python snippet that connects to a deployed model using the OpenAI client and ask candidates to fill in one placeholder value. Candidates need to understand the relationship between the Foundry project, the Azure OpenAI resource, and the model deployment, and know which identifier is used at each point in the connection code. Scenario questions describe a requirement to call a deployed model from an application and ask candidates to identify which information the application must include in its request.
Create and test a single-agent solution in the Foundry portal	True/False questions test the behaviour of the tool_choice parameter for agents in Microsoft Foundry Agent Service. Candidates need to understand what each possible value of this parameter causes the agent to do and be able to distinguish between the different options. Scenario questions describe a specific tool-calling requirement for an agent and ask candidates to select the correct tool_choice value that enforces that behaviour. Questions also test what must be configured to define an agent's role and behaviour, and candidates need to know which configuration component controls this versus other agent settings.
Lightweight client application for an agent	Scenario questions describe a requirement to use a Foundry capability and ask candidates to identify which Azure resource must be provisioned to enable it. Candidates need to distinguish between resource types that appear similar but serve different purposes within the Foundry ecosystem.
Text analysis application	Scenario questions describe a text processing task and ask candidates to select the correct text analysis technique from a list of named capabilities. Candidates need to know what each technique does and how to identify which one applies to a described requirement.
Respond to spoken prompts by using a deployed multimodal model (Voice Live)	True/False questions test what Voice Live is, what it produces, and how it differs from implementing speech capabilities separately. Multiple statements about the same feature are grouped together and test it from different angles, so candidates need a precise understanding of what Voice Live does as a single integrated capability.
Build a lightweight application by using Azure Speech in Foundry Tools	Scenario questions describe a speech processing requirement and ask candidates to select the correct Azure Speech capability. Candidates need to know the distinct use cases for each capability and understand the differences between processing modes — for example, the distinction between how pre-recorded audio and live audio streams are handled. Code-completion questions present a Python snippet using the Azure Speech SDK and ask candidates to fill in a placeholder value. Questions cover both speech synthesis and speech recognition scenarios, and candidates need to know which SDK class or method is appropriate for each context. Scenario questions also ask which SDK class is responsible for configuring the connection to the Azure Speech service, and candidates need to distinguish this from classes that handle audio input and output.
Interpret visual input in prompts by using a deployed multimodal model	Sentence-completion and multi-select questions test the formats in which images can be provided to a vision-enabled model via the OpenAI Responses API. Candidates need to know which formats are supported and which are not. True/False questions test specific constraints and capabilities around image input — including what is and is not possible in the Foundry playground and what is and is not required when including images in prompts. Questions test which content item type to include in a Responses API request when providing an image, and candidates need to know the correct field name from a list of plausible-looking options. Scenario questions describe a requirement involving both text and image content and ask candidates to identify the correct message structure. Candidates need to understand how content items are organised within a request and what structure produces the intended outcome. Questions also test where an application sends its requests after a vision-enabled model has been deployed in Foundry, and candidates need to distinguish the model endpoint from other Foundry resources.
Create new visual outputs by using generative models	Scenario questions describe a requirement to produce new images from text input and ask candidates to identify the correct AI workload type. Candidates need to distinguish image generation from other vision-related workloads where the goal is to analyse or extract information from an existing image.
Build a lightweight application that includes vision capabilities	Code-completion questions present a Python snippet using the OpenAI Responses API that sends a message containing both text and image content in a single request. Candidates are asked to fill in one or two placeholder values representing the content type fields, and need to know the correct type identifier for each content item. Scenario questions describe a requirement to include a local image as binary data in a request and ask candidates to identify the correct format for the image URL field. Candidates need to know which format is valid and be able to distinguish it from similar-looking but incorrect options.
Extract from documents and forms by using Azure Content Understanding in Foundry Tools	Scenario questions describe a content type and ask candidates to identify which analyzer type should be used to process it. Candidates need to know the correct analyzer type for each category of content supported by Azure Content Understanding. True/False questions test the capabilities and output characteristics of Azure Content Understanding, including what content types it supports and what format its results are returned in. Sentence-completion questions test what defines the fields to extract during an analysis operation, and candidates need to identify the correct term from a list of alternatives. Scenario questions describe a structured extraction requirement and ask candidates to select the correct tool from a list that includes other Azure services. Candidates need to know when Azure Content Understanding in Foundry Tools is the appropriate choice versus other extraction or processing approaches.
Extract from images by using Content Understanding	Questions test which analyzer type applies to image-based content. Candidates also need to understand the range of content types that Azure Content Understanding supports, as questions test whether candidates incorrectly assume the service is limited to a single content format.
Extract from audio and video by using Content Understanding	Scenario questions describe an audio file format or recording type and ask candidates to identify the correct analyzer type. Candidates need to know which analyzer handles audio content and be able to distinguish it from the analyzers used for other content types. Sentence-completion questions test which specific feature of Azure Content Understanding is responsible for converting audio to text during an analysis operation. Candidates need to distinguish this from other text-extraction technologies covered elsewhere in the exam.
Build a lightweight information extraction application using Content Understanding	Code-completion questions present a Python snippet using the Content Understanding SDK and ask candidates to fill in a placeholder value. Candidates need to understand the asynchronous pattern used to submit content and retrieve results, and know which method call belongs at each step. Scenario questions describe a workflow for submitting a document and retrieving extraction results, and ask candidates to select the correct sequence of SDK calls from a list of options that includes both valid and invalid approaches. Sentence-completion questions test how Azure Content Understanding processes submitted content — specifically whether processing is synchronous or asynchronous — and candidates need to select the correct term. Scenario questions test which Azure resource must be provisioned before Azure Content Understanding in Foundry Tools can be used, and candidates need to distinguish between resource types that appear related but serve different purposes.

← Domain 01: AI Concepts & Responsibilities

Common Questions

Domain 02 — Frequently Asked Questions

Topic-specific questions about what AI-901 tests in Domain 02 and how to approach the key implementation areas.

What question formats appear in Domain 02 of AI-901?

Domain 02 uses four question formats: scenario-based multiple choice (a situation is described and candidates choose the correct service, tool, or configuration); True/False (a statement about a specific behaviour or feature is presented and candidates confirm whether it is correct); sentence-completion (a sentence with a blank is provided and candidates choose the word or phrase that correctly fills it); and Python code-completion (a short code snippet is shown with one or two placeholder values that candidates must identify from a list of options). Code-completion questions are more common in Domain 02 than in Domain 01, so candidates should be comfortable reading short Python snippets.

How are system prompts tested in Domain 02?

System prompts are tested through True/False questions about what they do and do not control, sentence-completion questions about what must be configured to define an agent's behaviour, and multi-select questions about the purposes of instructions when prompting a generative AI model. Candidates need a precise understanding of what system prompts govern versus what is controlled by other configuration mechanisms in Foundry.

How is the Foundry SDK tested in Domain 02?

Code-completion questions present short Python snippets that connect an application to a deployed model and ask candidates to fill in placeholder values. Candidates need to understand the relationship between the different resources involved in a Foundry deployment and know which identifier is used at each point in the connection code. Scenario questions also ask what information a client application must include when calling a deployed model.

How is the tool_choice parameter tested for AI-901 agents?

The tool_choice parameter is tested through True/False questions about the behaviour each value produces, and through scenario questions where candidates must select the correct value to enforce a described tool-calling requirement. Candidates need to know what each possible value of this parameter causes the agent to do and be able to distinguish clearly between the options.

How is Azure Content Understanding tested in Domain 02?

Azure Content Understanding is tested across multiple question formats. True/False questions test the capabilities and limitations of the service. Sentence-completion questions test processing behaviour and what defines the extraction output. Scenario questions ask candidates to select the correct analyzer type for a given content format, choose between Content Understanding and other Azure services, and identify which Azure resource must be provisioned before the service can be used. Code-completion questions test the SDK pattern for submitting content and retrieving results.

How is image input to a vision-enabled model tested in Domain 02?

Image input is tested through sentence-completion and multi-select questions about the formats in which images can be provided via the OpenAI Responses API, True/False questions about constraints and capabilities in the Foundry playground, questions about the correct content item type to include in a request, and scenario questions about message structure when combining text and image content. Code-completion questions test the correct field values to use when sending image data from an application.

How is Azure Speech in Foundry Tools tested in Domain 02?

Azure Speech is tested through scenario questions that ask candidates to select the correct capability for a described requirement, code-completion questions involving the Azure Speech SDK where candidates fill in a placeholder class or method name, and questions that ask candidates to identify which SDK class is responsible for a specific part of the speech workflow. Candidates need to know the distinct capabilities within Azure Speech and understand how the SDK is structured across both synthesis and recognition scenarios.

Implement AI Solutions by Using Microsoft Foundry