Natural language is really messy… Could go through many variants on things. Then you get text to speech issues due to audio quality / accents… And you need an engine that can “best guess / best match” based on what it has or ask for clarification.
Similarly you can ask for TWO of a complex thing: I would like Two… meals, with, XXXX
An LLM can somewhat smooth over variances in language without having to have all possible variances known just the valid options and the raw input.
Mc Donald’s already has customer self serve kiosks and mobile apps with the full menu that limit you as to which items you can add or remove.
How did they screw this up and leave things open ended for the LLM?
IE why was the LLM not referencing a list of valid options with every request and then replying with what the possible options are. This is something LLMs are actually able to do fairly well, then layer on top the EXACT same HARD constraints they already have on the kiosk and mobile app to ensure orders are valid?
The self serve terminals and apps actually work well. I prefer using them over ordering at the counter.
So ya I am surprised they rolled this out so poorly.