The current AI wave gets framed as if it started in 2022. From my chair it didn't. I was shipping production NLP in 2017: an AI-powered dining reservation bot built on Azure Bot Framework, Node.js, and LUIS that interpreted guest intent across hundreds of venues.
The model behind the conversation matters less than getting the intent taxonomy right. If your bot can't classify what the user actually wants, no amount of model upgrade saves it. That was true with rule-based NLU, it stayed true with LSTMs, and it's still true with frontier LLMs.
The right answer is rarely "tune the model." It's usually "narrow the scope, ground the answer, fall back gracefully." The bot running on this site has a fixed knowledge brief and a single fallback sentence for everything outside that brief. The fallback is the feature.
A response under 800ms feels conversational. Over two seconds, the conversation breaks. That budget dictates retrieval strategy, model choice, and streaming approach more than any benchmark chart.
LLMs raised the ceiling on what these systems can do. The floor still needs the same engineering discipline: intents, grounding, latency, fallbacks. The model is the engine. Everything else is still the product.
Treat the model as the engine, not the product.