Artificial intelligence is not a single workload. An organization might predict demand, detect sentiment, transcribe speech, recognize an object, extract a field or generate an explanation. The same application can contain several such capabilities, but each has a different relationship between input, output, authority and evaluation. In Microsoft AI-901, choosing an appropriate workload matters more than recognizing the name of every historical Azure AI service.<\/p>\n
Suppose a small municipality wants a digital assistant for public service requests. Residents submit written messages, photographs of damaged roads, voice messages and scanned forms. Some requests need simple classification; others require careful extraction or a human decision. The correct architecture does not route every input to a single open-ended chatbot. Begin with what has to be recognized, transformed or produced, then choose the smallest toolchain that satisfies the requirement.<\/p>\n
Traditional machine-learning workloads often predict a label or numeric value from observed features. For example, a model can predict whether a service ticket is likely to miss its target resolution time. A generative model creates content such as a summary, response, image or explanation based on its inputs and learned patterns. A prediction about road repair volume is different from a narrative explaining why a repair backlog grew. The output type determines what a meaningful error looks like.<\/p>\n
For classification, examine false positives and false negatives and whether the label distribution changes. For generative answers, inspect factual grounding, completeness, tone and unsupported claims. An apparently elegant summary can omit a safety-critical detail. Do not replace a high-stakes numeric forecast with confident free text unless the underlying evidence and calculation remain accessible.<\/p>\n
An agent can use a model to interpret a goal, choose a supported tool or coordinate a bounded sequence of steps. For the municipality, an agent might ask a resident for a missing location, create a draft service request and retrieve its status. The model’s ability to draft a message does not grant permission to submit or cancel the request. An agent needs explicit tool contracts and authorization for every meaningful action.<\/p>\n
A static FAQ may need only retrieval and response generation. A workflow that updates an official record may need an agent or deterministic orchestration with review, validation and recovery. The single-agent Foundry walkthrough<\/a> examines that difference. When a model returns a plausible \u201cticket created\u201d message without a confirmation from the ticketing system, the workflow is not complete.<\/p>\n Key-phrase extraction identifies important topics in a message. Entity recognition identifies mentions such as names, organizations, dates or locations. Sentiment analysis estimates an opinion’s polarity, and summarization condenses a longer passage. These are different outcomes. A highly negative tone does not by itself identify whether a customer requested a refund or reported an equipment hazard.<\/p>\n Choose representative test messages and compare the extracted result with a human-reviewed expectation. A location entity can still be the wrong place for a work order; the application must resolve it against actual addresses and business records. The text-analysis practice cases<\/a> show how to combine these capabilities without treating an extracted field as automatically verified.<\/p>\n Speech recognition converts audio to text and can support voice-driven service requests. Speech synthesis produces spoken output from text. A conversational system may use both and pass the recognized text to a language model. It matters that the stages are separate: a transcription error can contaminate the model’s reasoning, while a synthesis issue can make a correct answer hard to understand. A multimodal model may accept audio directly, but that support must be checked for the selected deployment.<\/p>\n For a road hazard hotline, test unusual street names, background noise and emergency vocabulary. Let the caller verify an address before creating a record. Where communication is important, provide a transcript or alternate input channel. The speech recognition and multimodal audio<\/a> explains why recognition accuracy, synthesis voice quality and the application’s final business state require different checks.<\/p>\n Computer vision interprets visual inputs: objects, scene content, text in an image or relationships relevant to a task. An image-generation model produces new visual output from instructions. If a resident photographs a damaged sidewalk, the first question may be whether the image contains a sidewalk and whether an issue is visible; creative generation would be the wrong core capability. If the communications team needs an illustration of a proposed park redesign, a generator may be appropriate with suitable disclosures and rights checks.<\/p>\n A vision model that can describe \u201ca damaged surface\u201d does not necessarily provide a calibrated severity measurement. A repair priority should be based on approved criteria and, where needed, inspection. Distinguish visual input analysis<\/a> from image-generation controls<\/a> when studying the exam.<\/p>\n A scanned form often needs exact fields and source references rather than a broad summary. Information-extraction tools can identify names, dates, amounts, tables and other structured elements. Document extraction differs from OCR alone: recognizing characters is useful, but assigning each value to the correct field and distinguishing a subtotal from a final amount requires additional interpretation and validation. Documents can also contain multiple pages, signatures and handwritten notes.<\/p>\n Azure Content Understanding in Foundry Tools can process documents, images, audio and video through configured analyzers. The document and form extraction<\/a> explains how schemas and confidence evidence support validation. The municipality could use extraction to prefill an application while leaving the resident to confirm the required details.<\/p>\n Make a small decision table with the source type, desired result, likely technique and key failure condition. A voice message becomes text through speech recognition; the transcript can be classified; a photo can be examined by a vision model; a form can be extracted to fields; and a response can be composed by a generative model. The sequence may be useful, but each stage introduces errors and privacy obligations that must be tested.<\/p>\nText analysis extracts insights rather than inventing answers<\/h2>\n
Speech has recognition and synthesis paths<\/h2>\n
Computer vision is not always image generation<\/h2>\n
Information extraction produces structured records<\/h2>\n
Select one workload at a time before composing a system<\/h2>\n