AI-900 Cheat Sheet

Key concepts

• ✓Matching scenarios to Azure AI workloads: vision, language, document intelligence, generative AI
• ✓Identifying responsible AI principles: fairness, reliability and safety, privacy, inclusiveness, transparency, accountability
• ✓Recognizing Azure services like Azure AI Vision, Azure AI Language, and Azure OpenAI
• ✓Distinguishing AI workload categories from general data, analytics, or infrastructure services

Watch out for

• ⚠Confusing Azure AI Language with Azure AI Vision when a scenario involves extracting meaning from text versus images.
• ⚠Mixing up transparency and accountability: transparency means explaining AI decisions, accountability means people remain responsible.
• ⚠Assuming any automation or analytics feature counts as AI, when the exam expects a specific cognitive workload category.

Key concepts

• ✓Selecting regression, classification, or clustering for a given business scenario
• ✓Identifying Azure Machine Learning workspace assets: datasets, compute targets, and environments
• ✓Using Automated Machine Learning to train and deploy models with minimal code
• ✓Building ML pipelines with Azure Machine Learning designer's drag-and-drop modules

Watch out for

• ⚠Confusing clustering (unsupervised, no labels) with classification (supervised, labeled categories), which leads to wrong model choice on scenario questions.
• ⚠Assuming Automated Machine Learning removes all data preparation; it still needs clean, correctly formatted tabular data with a defined target column.
• ⚠Mixing up Azure Machine Learning compute targets: compute instances for authoring versus compute clusters for scalable training jobs.

Key concepts

• ✓Choosing between Azure AI Vision, Face, and Document Intelligence for a given scenario
• ✓Distinguishing image classification, object detection, and semantic segmentation capabilities
• ✓Recognizing OCR, Read API, and handwriting extraction use cases in Azure AI Vision
• ✓Identifying facial detection versus facial recognition limits and responsible AI constraints

Watch out for

• ⚠Assuming Azure AI Vision can identify a specific person by name; that is Face service recognition, which is restricted.
• ⚠Confusing object detection, which returns bounding boxes, with image classification, which returns only category labels.
• ⚠Treating Document Intelligence as general image analysis rather than form, invoice, and receipt extraction.

Key concepts

• ✓Static NAT, dynamic NAT and PAT behaviour.
• ✓Inside local, inside global, outside local and outside global address meanings.
• ✓How NAT affects connectivity between private and public destinations.
• ✓How to troubleshoot NAT rules, ACL matches and interface direction.

Watch out for

• ⚠PAT allows many inside hosts to share one public address by using port numbers.
• ⚠NAT rules depend on correct inside and outside interface configuration.
• ⚠The ACL used for NAT identifies traffic to translate; it is not a security ACL.
• ⚠Static NAT maps one private address to one public address; PAT overloads translations.

Key concepts

• ✓Distinguishing generative AI, which creates new content, from discriminative models that classify or predict labels
• ✓Identifying Azure OpenAI Service capabilities including GPT models, DALL-E image generation, and embeddings
• ✓Recognizing prompt engineering basics: system, user, and assistant messages plus completion outputs
• ✓Applying Microsoft's responsible AI principles to generative workloads, including transparency, safety filters, and content moderation

Watch out for

• ⚠Assuming Azure OpenAI Service hosts only OpenAI models, ignoring that it also offers Microsoft and third-party models through Azure AI Foundry
• ⚠Confusing a model deployment name with the underlying base model name, which breaks API calls and endpoint configuration
• ⚠Treating hallucination as a bug fixed by retraining, rather than an inherent limitation mitigated with grounding, retrieval, and prompt design