I keep seeing the same mistake in n8n workflows: people default to the AI Agent node for everything. In this video, I walk through the small set of AI nodes that covers the majority of real-world automation work—so you stop building “Swiss Army knife” flows when a purpose-built node will be cheaper, more reliable, and easier to maintain. I show where the AI Agent actually makes sense (tools, memory, chat-style workflows, output parsers), and why you should almost always set a system message if you care about accuracy. Then I break down the specialized nodes: Basic LLM Chain (use it when you don’t need tools/memory), Text Classifier (great for routing support tickets and requests), Sentiment Analysis (brand/social monitoring), Information Extractor (turn messy text into structured fields like name + salary), Summarization Chain (with chunking + map-reduce/refine options), and Q&A Chain for RAG (vector store + embeddings + retrieval). I also cover Guardrails for security (PII, jailbreaks, topical alignment), Evaluation nodes to score workflow outputs, why I’m cautious with AI Transform, and model-specific nodes (Anthropic, Gemini, OpenAI). Finally, I explain why mastering the HTTP Request node is non-negotiable when n8n doesn’t have the integration you need.