NLG is the layer that produces the words the agent says. In legacy CX systems, NLG was a template engine with slots ("Hi {name}, your appointment is confirmed for {date}"). In LLM-driven systems, NLG is the model itself generating the reply token by token.
The shift sounds like an improvement and mostly is, but it introduces failure modes templates did not have. Templates cannot hallucinate; LLM-generated replies can. Templates cannot drift from brand voice; LLM replies need careful prompting and evaluation to stay consistent. Templates cannot say something rude when stressed; LLMs occasionally can.
Production-grade NLG combines the strengths. The agent generates freely on most turns, but high-stakes confirmations (booking details, policy quotes, escalation language) fall back to typed templates with verified values. The customer cannot tell which mode is which; the engineering team has confidence both work.
Vorel uses LLM NLG for conversational turns and templated NLG for confirmations, policy quotes, and the closing summary. Mode selection is per turn-type, not per call.