What you must be able to do. Match a stated concern to the correct responsible-AI principle, reason about how generative models work and how you select and configure one, and map a scenario to the right AI workload.
In one sentenceThe conceptual half: the six responsible-AI principles, how generative models work and how you choose and configure them, and matching a scenario to the correct AI workload.
Recall check: answer these from memory first
- Name the six Microsoft responsible-AI principles and give the one concern each addresses.
- State in one line the difference between transparency and accountability as responsible-AI principles.
- Given a scenario, say whether it is generative, agentic, text analysis, speech, computer vision, or information extraction, and why.
What it tests. The six Microsoft responsible-AI principles and the consideration each addresses; how generative models work at a conceptual level, including tokens and next-token prediction, plus selecting a model by capability and the deployment and configuration parameters such as temperature that shape its behaviour; and identifying which AI workload a scenario needs, across generative and agentic AI, text analysis, speech, computer vision, and information extraction, keeping the text-analysis techniques distinct.
How to study it. Learn the six principles as concern-to-principle mappings rather than definitions, and drill the transparency-versus-accountability pair until it is automatic because it is the most-tested confusion. For models, anchor the sampling parameters to outcomes you can reason about: a low temperature gives reproducible, on-message output. For workloads, practise placing a scenario into exactly one category, and be strict that an agent that plans and takes actions is agentic, not generative. Keep keyword extraction, entity detection, sentiment analysis, and summarization separate in your mind.
Easy to confuse
- Transparency versus accountability. Transparency is making the system and its decisions understandable to the people affected, including its limits; accountability is people and governance remaining answerable for how the system behaves. If the concern is understanding a decision it is transparency; if it is who is responsible and what oversight exists it is accountability.
- Agentic AI versus generative AI. Generative AI produces content such as text or images in response to a prompt; agentic AI plans and takes actions, calling tools or knowledge, to complete a task towards a goal. If the scenario describes taking steps and using tools rather than just producing an answer, it is agentic.
- Selecting a model by capability. A language model handles text, a multimodal model accepts images or audio alongside text, an embedding model turns content into vectors for similarity, and an image-generation model creates pictures. The exam states a capability constraint and expects the one model class that meets it, so match the stated need rather than defaulting to a general chat model.
Worked example from the AI-901 bank
lock_openFree sampleIdentify AI concepts and capabilitieseasy
A bank deploys a model that declines some loan applications. The team wants applicants and reviewers to be able to understand which factors drove each decision and how the system reaches its conclusions. Which Microsoft responsible AI principle does this goal address?
- ATransparencycheck_circle Correct
- BFairness
- CAccountability
- DReliability and safety
Transparency means making an AI system and the factors behind its decisions understandable to the people it affects. The requirement is that people can understand how the system works and why it produced a particular outcome, which is precisely what the transparency principle is intended to deliver.
Why A is correct: Transparency is the principle of making an AI system understandable, so that people can see how it works and which factors influenced a given decision.
Why B is wrong: Loan decisions are a classic setting for bias concerns, which makes fairness tempting, but fairness is about giving comparable groups comparable treatment, not about making the reasoning behind a decision understandable.
Why C is wrong: Being answerable for decisions sounds close, but accountability is about people and governance remaining responsible for the system, not about the system itself being intelligible to those affected.
Why D is wrong: Reliability and safety covers consistent, safe operation under varied conditions, so it is tempting for a high-stakes banking use, but it does not concern explaining how a decision was reached.
What you must be able to do. Choose the correct Microsoft Foundry surface and capability for a described build, across generative and agentic apps, text and speech, vision and image generation, and information extraction.
In one sentenceThe larger, more technical half: building with Microsoft Foundry across generative and agentic apps, text and speech, vision, and information extraction, using the right surface and capability each time.
Recall check: answer these from memory first
- Say when a build needs the Foundry portal and when it needs the Foundry SDK.
- Distinguish a multimodal model interpreting an image from a generative image model creating one.
- Name the Foundry capability for extracting structured fields from documents, and say why a general vision model is the wrong choice.
What it tests. Implementing AI solutions with Microsoft Foundry: writing prompts and deploying and interacting with a model in the Foundry portal, building with the Foundry SDK, and creating and testing a single-agent solution; text analysis and responding to spoken prompts, plus Azure Speech in Foundry Tools for speech-to-text and text-to-speech; interpreting visual input with a multimodal model versus generating new images; and extracting structured information from documents, images, audio, and video with Azure Content Understanding in Foundry Tools.
How to study it. This domain carries the larger weight, so spend the most time here. Build a mental map of Foundry: the portal for deploying and testing visually, the SDK for building an application in code, and the named capabilities (Azure Speech, Azure Content Understanding) that sit in Foundry Tools. Practise the single-best decisions the exam leans on: portal versus SDK, a multimodal model that interprets an existing image versus a generative model that creates a new one, and Azure Content Understanding versus a general vision model for structured field extraction. Use only current names in your notes so a retired one stands out as wrong.
Easy to confuse
- Foundry portal versus Foundry SDK. The portal is the visual surface for deploying a model, testing prompts, and creating and testing an agent; the SDK is the code path for building a lightweight client application. If the scenario is building an app in code it is the SDK; if it is deploying or testing visually it is the portal.
- Azure Content Understanding versus a general vision model. Azure Content Understanding in Foundry Tools is purpose-built to extract consistent structured fields from documents, images, audio, and video; a general multimodal vision model can read an image but is not built for repeatable structured output. A requirement to pull the same named fields from many documents points to Content Understanding.
- A multimodal model accepting speech versus Azure Speech. A deployed multimodal model can take a spoken prompt directly as input; Azure Speech in Foundry Tools is the dedicated capability for speech-to-text and text-to-speech. If the task is a standalone recognition or synthesis feature it is Azure Speech; if it is feeding spoken input into a multimodal model it is the model itself.
Worked example from the AI-901 bank
lock_openFree sampleImplement AI solutions by using Microsoft Foundryhard
A developer has finished prototyping a chat experience in the Microsoft Foundry portal and now needs to embed the same deployed model inside a C# web service so that inference requests are issued programmatically at runtime. Which Foundry surface is designed for this?
- AThe Foundry SDK, which exposes client libraries so an application can authenticate to and call the deployed model from its own code.check_circle Correct
- BThe Foundry portal, which is a browser-based workspace for building and testing projects interactively rather than for embedding calls in application code.
- CAzure Content Understanding, which extracts fields from documents and media rather than providing programmatic access to a deployed chat model.
- DThe model catalogue, which lists available models to browse and deploy but does not itself issue runtime inference requests from application code.
The Foundry SDK is the surface an application uses to call a deployed model programmatically, whereas the portal is for interactive work. Embedding a deployed model in application code requires client libraries that handle authentication and inference calls at runtime; the Foundry SDK provides exactly these, while the portal is an interactive design and testing environment.
Why A is correct: The Foundry SDK provides client libraries for languages such as C# and Python, letting application code authenticate and send inference requests to a deployed model at runtime, which is exactly what embedding the model in a web service requires.
Why B is wrong: The portal is tempting because that is where the prototyping happened, but it is an interactive browser workspace for building and testing, not a mechanism for issuing calls from within a compiled application.
Why C is wrong: Azure Content Understanding is an information-extraction capability for documents, images, audio, and video, so it does not give an application programmatic access to a deployed conversational model.
Why D is wrong: The model catalogue helps you find and deploy a model, but it is a discovery surface, not the runtime client an application uses to call the model once deployed.