"The exam doesn’t test you on what GPT-4 can do — it tests you on how to configure it. Know your deployment types, temperature vs max tokens, and when to use system messages vs few-shot examples. RAG questions come up a lot more than you’d expect."

"The new AI-102 outline calls agents out separately. Treat an agent as a hosted reasoning loop with instructions, tools, memory/conversation state, and governance. The exam is likely to ask when to use a prompt agent, a workflow agent, or Microsoft Agent Framework for more complex orchestration."

"Computer vision was easier than I expected — most questions boil down to: use standard Azure AI Vision if it’s a common task (tags, objects, OCR), use Custom Vision if you have your own categories to train on. Don’t overthink it. OCR Read API questions were very common though."

"NLP questions are usually service-selection questions. Language analyses existing text, Speech handles audio input/output and SSML, Translator handles text/document/speech translation, CLU identifies intent, and custom question answering returns answers from a curated knowledge base."

"This domain is two related pipelines: Azure AI Search for knowledge mining and retrieval, then Document Intelligence / Content Understanding for turning files into structured information. Know the nouns: data source, indexer, skillset, index, knowledge store, semantic ranker, vector store, prebuilt model, custom model, analyzer."