A common failure mode in AI engineering is over-engineering: reaching for an LLM or vector database when a simpler solution would work better. You are asked to build a customer support routing system that classifies incoming tickets into 15 categories and routes them to the right team. Walk through your decision process: when does this need an LLM, when does fine-tuned BERT or a classifier work better, when does simple keyword matching or business rules suffice? Describe a framework for making this decision on any given AI engineering problem.