You give them a list of intents, a list of utterances that should invoke those intents and “slots” that those intents need to fulfill the intent.
An utterance would be “I want to go home”.
It would match an utterance to “I want to go $location”. That is matched to an intent “get directions” and then you route it to the correct subsystem. If you don’t have all of the slots (ie they don’t mention the location), you keep asking questions until all of the slots are filled.
Any first year hobbyist programmer can replicate the natural language processing of Siri.
Yes the harder part is converting speech to text.
This is what a more intelligent LLM based system could do.
This reminds me of the first "conversation" I had with Emacs's ELIZA. I have it saved up since reading it always makes me smile.
AI asks me a ton of questions to learn about me for better targeted advertising.