Retrieval-augmented generation gets treated, in a lot of pitches, as if it's a complete AI strategy on its own. It isn't — it's a technique for grounding a language model's answers in a specific set of documents, and it's genuinely useful for exactly that.
What it doesn't solve on its own: what happens when the retrieved documents contradict each other, how the system behaves when nothing relevant is found, or how confident the answer should sound relative to how confident the retrieval actually was.
The projects that work well treat RAG as one component in a system that also includes clear fallback behavior, some notion of confidence, and a defined path to a human when the system is out of its depth.