But what Jeff was facing wasn't an LLM. As far as I saw it was pattern matching with hardcoded responses. Hence it had a hardcoded response about WiFi, after sa

But what Jeff was facing wasn't an LLM. As far as I saw it was pattern matching with hardcoded responses. Hence it had a hardcoded response about WiFi, after saying "Not WiFi". ChatGPT etc wouldn't make such a mistake. And the final human also didn't understand the request correctly. ChatGPT wouldn't respond with "I see you have a question about your Wi-Fi" after "Slow Internet, not slow WiFi". I think this pattern matching was more like "automated questionnaire filtering" that you were suggesting. It's just it's very hard to do this filtering, to account for differences between "not WiFi" vs "WiFi", while it's easy with LLMs.