NLU is the historical category name for the layer that takes language input and produces structured meaning, the layer below dialogue management and above raw text. It covered intent classification, entity extraction, sentiment, coreference, anaphora, and a dozen other sub-problems that each had their own model.
In the LLM era, most of these sub-problems are handled implicitly by the same model that produces the response. There is no separate intent classifier; the LLM reads the utterance, infers the intent, and acts on it in one pass. NLU as a separate layer is fading.
It still matters as a concept for two reasons. First, the failure modes of NLU (misunderstood intent, missed entity) still exist, they just live inside the LLM now and need different debugging tools. Second, some production systems keep a discrete NLU layer for low-latency hot paths where running a full LLM turn would be too slow, the classifier returns a routing decision in 50 ms while the LLM is still warming up.
Vorel runs LLM-native NLU on the main turn loop, with a small discrete classifier for the first-utterance routing decision where latency is most caller-perceptible.