Microsoft Azure AI Fundamentals (AI-901) — Practice Questions and Mock Exam

Two facts combine into the answer. First, image understanding is delivered by vision-enabled chat models, large multimodal models that analyze images and answer questions about them in text. Second, you use them by calling the chat completions interface, which is designed for multi-turn conversations. So the image arrives in a user message, follow-up questions arrive as later turns, and the model reasons over the picture together with everything said so far. Asking about the jacket's color and then whether it suits rain is simply two turns of one conversation.

The distractors each do something real, but not this. An image generation model produces pictures from text, which is the opposite direction of travel. The tagging feature of the image analysis service returns labels and confidence scores for a single image and holds no conversation, so a follow-up question has nowhere to go. The embeddings interface turns content into vectors for similarity search, which is useful for finding a similar product and useless for answering a question in words.

Exam tip: when a scenario mentions follow-up questions or context carried across turns, the answer lives on the chat surface. Feature-style vision services return a fixed result set per call and cannot hold a dialogue.

Start by classifying the input, because that is what separates these four services. The article text is written text, and the job is turning written text in one language into written text in another. That is neural machine translation, which is the Translator service: it runs in real time as well as in batch, covers a long list of languages, and also handles transliteration and language detection.

The near-miss is the Speech service. It does translate, but it translates spoken audio, through speech-to-text and speech translation pipelines. Nothing in this scenario involves audio, so reaching for it would mean paying for machinery the workload never uses. Content Safety inspects text and images for harmful material and returns severity scores; it never produces a translation. Vision extracts information from images, which is the wrong modality entirely.

Exam tip: on the workload-identification questions, match the input modality first, whether that is text, audio, image or video, and only then choose among the services that accept it. That single habit resolves most service-selection items on this exam.

A provisioned throughput unit is a unit of capacity, not a unit of money, of text, or of configuration. When you choose a provisioned deployment instead of paying per token you are reserving a fixed slice of model processing capacity and holding it for your workload, and the number of units you allocate is how you size that slice. Capacity is granted as quota, and the same quota can be spent on any supported model, which is why the unit is deliberately generic rather than tied to one model.

The distractors each borrow a word from somewhere else on the platform. A fixed charge per API call describes a billing model, and provisioned deployments bill per unit per hour rather than per call. The number of tokens allowed in a single prompt is the context window, a property of the model. The maximum number of stop sequences is a request parameter and has nothing to do with capacity.

Exam tip: reserved-capacity questions turn on predictable throughput and consistent latency. A scenario with steady, high-volume traffic and a need for stable latency points at provisioned throughput, while bursty or experimental workloads point back to paying per token.

Three properties in the scenario point at the same workload type: the application plans a task over multiple steps, it calls external tools and APIs to fetch live data, and it decides what to do next with little human intervention. That combination is agentic AI. A managed agent runtime is built around exactly those pieces, managing the conversation, the tool calls and the agent lifecycle, while tools reach outside through a catalog, the model context protocol and OpenAPI definitions.

The distractors are all narrower, single-purpose workloads that produce one output from one input. Anomaly detection scores data points against expected behavior and raises a flag; it neither plans nor acts. Optical character recognition extracts text from an image. Image classification assigns a label to a picture. None of them calls a tool, carries state across steps, or decides what to do next.

Exam tip: the giveaway words for agentic workloads are autonomous, multi-step, tool or API calling, and goal. When a scenario has all four, no single-task perception or prediction workload can be the answer.