Free AI-901 practice test

Azure AI Language bundles core text NLP features: sentiment, key phrases, language detection, and entity recognition. Azure AI Vision (A) analyzes images, Document Intelligence (B) extracts form fields, and Face (C) analyzes faces — none is the general text-analysis service.

Transparency means people can understand how the system works and its limitations; accountability means humans remain responsible for the system and its outcomes, including governance and oversight. Inclusiveness (E) is about serving diverse users. Elasticity (A) and scalability (B) are cloud-infrastructure concepts, not responsible AI principles.

Treating people inequitably based on a characteristic like gender is a failure of fairness — the principle that AI systems should treat all people fairly and avoid bias. Reliability and safety (A) is about consistent, safe operation. Transparency (C) is about understandability. Accountability (D) is about human responsibility for the system.

Transcribing calls uses speech-to-text, and judging caller satisfaction from the transcript uses sentiment analysis — a common NLP pipeline. Text-to-speech (A) generates audio (not needed here), and object detection (B) and OCR (C) are computer vision tasks irrelevant to analyzing call audio and text.

Detecting and locating multiple objects with bounding boxes is object detection. Image classification (D) gives one label for the whole image without locations. Sentiment analysis (B) and text-to-speech (C) are NLP/speech tasks, not computer vision.

Azure AI Face specializes in detecting faces and analyzing facial attributes. Document Intelligence (B) extracts data from forms, Translator (C) translates text, and Content Safety (D) moderates content — none is the dedicated face-detection service.

Azure AI Vision offers ready-to-use image analysis, tagging, OCR, and captioning through an API. Azure AI Language (C) handles text, Azure AI Speech (D) handles audio, and Azure Machine Learning designer (B) is for building custom models — not prebuilt vision features.

Interpreting the content of images — identifying and locating objects in a photo — is a computer vision workload. Natural language processing (C) works with text and speech. Anomaly detection (A) finds outliers in numeric data. Knowledge mining (D) extracts information from documents, but the input here is images, not text.

Determining which language a piece of text is written in is language detection. Named entity recognition (A) finds people/places/organizations, sentiment analysis (C) scores emotion, and speech recognition (B) transcribes audio — none identifies the language of text.

Azure Speech's text-to-speech offers neural voices that sound natural and can be selected/configured for language and style to match a desired brand voice. Audio has no 'pixel resolution', switching to computer vision is irrelevant to audio output, and disabling text-to-speech removes the very feature being requested.

The portal exposes the run or thread details for an agent conversation, including which tool calls were made and with what arguments, so you can debug behavior. Cost Management, regional availability, and IAM role pages are unrelated administrative views that do not show per-turn tool call details.

An agent's instructions function like a system prompt: they define its role and behavioral constraints for the entire session, not just one exchange. They are not a transcript log, they do not control billing tier, and they are not a directory of other agents.

Content Understanding is designed to work across modalities present in video — audio for spoken dialogue and visual frames for on-screen text — with the analyzer configuration determining which are extracted. Claiming it only handles audio, only handles visuals, or does not support video at all understates its actual multimodal extraction capability.

Creating new images from a text description is image generation using a generative model — exactly the described use case. Text analysis, speech-to-text, and document form extraction are all different capabilities that operate on existing text, audio, or documents rather than creating new images.

Content Understanding's strength is extracting and recognizing text (including handwriting) and structured fields from images; interpreting the semantic meaning of a free-form diagram is a harder, more limited task than text recognition. It does not refuse images containing diagrams outright, it does not produce audio narration, and it certainly does not substitute unrelated stock imagery.

Maintaining the image within the conversation's context or thread lets subsequent messages reference it without re-uploading, similar to how text history is preserved in chat. Requiring the user to describe it from memory defeats the purpose of visual understanding, deleting the image blocks any follow-up, and forcing new unrelated images each time breaks the intended workflow.

This is a vision-input scenario: the multimodal model reads the actual image content alongside the text instruction to produce a response grounded in what's visually present. It is not generating a new image, it is not restricted to text-only input, and it is not converting the image to audio.

Content Understanding's core strength on images is recognizing and extracting text-based fields per the analyzer's configuration; whether a barcode is decoded depends on the analyzer setup rather than being an automatic side effect. It does not place orders, does not refuse images containing barcodes, and does not generate spoken audio from a barcode.

For large offline batches, applications should use asynchronous/batch-style processing — submitting jobs and polling for results — rather than requiring interactive, one-at-a-time handling, which does not scale. Manually triggering each call in the portal is impractical at volume, redeploying a new model resource per document is wasteful and unnecessary, and emailing raw receipts to customers does not fulfill the extraction/reporting requirement.

Content Understanding extraction is driven by a defined analyzer or schema that specifies which fields matter for a given document type; that configuration determines what is returned. It does not indiscriminately return every conceivable field, Azure region affects hosting/latency not extracted fields, and browser language has no bearing on server-side extraction results.