AI Workloads & Considerations

Coming in you need only everyday familiarity with what AI can do; this domain leaves you able to do two things: classify a scenario into a workload category, and name the responsible-AI principle a scenario raises. Domain 1 of AI-900 has two halves that the exam keeps separate. The first half, covered by the AI Workload Types subtopic, asks you to read a business scenario and classify it into one of five workload categories: machine learning, computer vision, natural language processing (NLP), document intelligence / knowledge mining, and generative AI. The second half, covered by Responsible AI Principles, asks you to read a scenario about how an AI system affects people and name which of Microsoft's six responsible-AI principles it illustrates. The first lens is about capability (which Azure service family solves this problem); the second is an overlay that applies to every workload regardless of category. A reliable signal for which subtopic a question targets: if it names a task or input/output ("extract fields from invoices," "summarize a document," "detect objects in a photo"), it is a workload-classification question; if it names a value or harm ("the model disadvantages one gender," "users should know how a decision was made"), it is a responsible-AI question.

The fastest discriminator across the whole workload half of this domain is to ask what goes in and what comes out, because each category has a distinct input/output signature. Tabular data in and a predicted score or class out is machine learning; an image or video in and labels, bounding boxes, or extracted text out is computer vision; unstructured text or audio in and structure (sentiment, entities, key phrases, a translation) out is NLP, and the exam files speech-to-text, text-to-speech, and speech translation under the NLP family. Semi-structured documents in and structured fields or a searchable index out is document intelligence / knowledge mining; a natural-language prompt in and brand-new original content (text, code, an image) out is generative AI. Once classified, AI-900 expects you to name the primary Azure service: custom ML → Azure Machine Learning, computer vision → Azure AI Vision, NLP → Azure AI Language and Azure AI Speech, document processing → Azure AI Document Intelligence, knowledge mining → Azure AI Search, and generative AI → Azure OpenAI within Azure AI Foundry.

Microsoft's Responsible AI Standard defines exactly six principles a system should satisfy simultaneously: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. The exam's recurring pattern is to give a short scenario and ask which one it illustrates, so the synthesis skill is knowing the discriminators between the easily confused pairs. Microsoft groups them into two memory pairings: fairness and inclusiveness are both about who the system serves (treat similar people alike; design for everyone), while transparency and accountability are both about human oversight (explain the decision; keep a human responsible). The remaining contrast is reliability/safety (how the system behaves under odd inputs) vs privacy/security (how data is protected and lawfully governed). The single most common AI-900 trap offers "choose the largest or most-accurate model and trust it" as the responsible answer. It is wrong because accuracy alone ignores fairness, safety, transparency, and human control. Critically, these six principles apply on top of whichever workload category you identified in the first half of the domain; a fairness or transparency obligation attaches to a vision model, an NLP model, and a generative model alike.