AI-901 asks candidates to recognize what different AI models and workloads are good at before they learn how to assemble solutions in Microsoft Foundry. The current blueprint covers generative and agentic AI, text analysis, speech, computer vision, information extraction, model capability, deployment options, and configuration parameters.<\/p>\n
The practical skill is classification: given a business problem, identify the workload and model capability that fits it. A model should be selected because its strengths match the task, not because it is the newest or largest option available.<\/p>\n
Predictive models estimate a class, value, or probability from input data. Generative models create new content such as text, images, or code based on learned patterns and the prompt context. Agentic systems add planning or tool use so the AI can take a sequence of steps toward a goal.<\/p>\n
These categories can overlap inside one application. A support agent may classify a request, retrieve information, generate a response, and call a business tool. AI-901 scenarios are easier when you identify which capability is required at each step.<\/p>\n
Model selection involves more than raw quality. Teams may need to balance reasoning ability, modality support, latency, cost, context size, deployment availability, and safety requirements. A small model may be ideal for a high-volume classification task, while a more capable model may be justified for complex reasoning.<\/p>\n
The right answer depends on the workload. Overprovisioning a model can increase cost and latency without meaningful quality improvement. Underprovisioning can create poor output or require complicated compensating logic.<\/p>\n
AI models can be exposed through different deployment options depending on the platform and workload. Configuration parameters influence behavior: for example, randomness controls can affect how deterministic or varied a generative response becomes, while token limits influence response length and resource use.<\/p>\n
AI-901 does not require deep mathematical derivation, but candidates should understand that model behavior is partly shaped by configuration. A prompt, model choice, and deployment configuration work together.<\/p>\n
Text analysis includes tasks such as keyword extraction, entity detection, sentiment analysis, and summarization. These workloads turn unstructured language into structured signals or shorter representations that an application can use.<\/p>\n
A common exam mistake is to choose a generative model for every language problem. Traditional language-analysis capabilities can be more appropriate when the requirement is structured extraction rather than open-ended generation.<\/p>\n
Speech recognition converts spoken audio into text or structured understanding, while speech synthesis generates spoken output from text. Translation and multimodal models can extend the experience, but the underlying requirement still matters.<\/p>\n
If an application needs to transcribe a call, the core workload is speech recognition. If it needs to read a response aloud, it is speech synthesis. If it needs to answer a spoken prompt with context from other data, several capabilities may be combined.<\/p>\n
Computer vision interprets visual input, such as identifying content in images or understanding visual prompts. Image-generation models create new visual outputs from instructions. Multimodal models can combine vision and language so an application can reason about an image and answer in text.<\/p>\n
For AI-901, distinguish understanding from generation. A system that reads a receipt or describes a photo is interpreting visual input; a system that creates a new illustration is generating visual output.<\/p>\n
Information extraction turns documents, forms, images, audio, or video into structured fields and entities. Azure Content Understanding in Foundry Tools is central to the current AI-901 objectives because organizations often need to transform messy business content into reliable data.<\/p>\n
This is different from generic summarization. A summary captures meaning in prose, while extraction aims to recover specific information that downstream systems can validate, store, or process.<\/p>\n
These foundational distinctions become more important as candidates move toward engineering roles. The Azure AI engineering career path<\/a> increasingly requires combining language, vision, retrieval, agents, and data services into one system rather than treating each capability in isolation.<\/p>\n