Siri is still built on the same ML model-based system, capable of learning, but from what this article says, the original engineers (pre-Apple?) created a poor design for the software and database, and Apple has had a hard time maintaining that.
There's two traditional models for speech recognition; command/control and dictation.
Command and control has a limited set of words and expected sentence structures, but as result has a much better clue at 'guessing' what you might want. It is like the old Palm Pilot - don't try to recognize cursive, get the user to jot things the way you want (Graffiti).
Dictation is freeform, but it has zero clue at what words mean or what sentences are. More flexible, much more error prone.
Siri is a hybrid of the two. A fixed dictionary, but it is per user with things like contacts influencing acceptable commands, and the ability to use giant databases such as all album titles in Apple Music. More flexibility in sentence structure, and the ability to switch to dictation mid-command, e.g. "text Steven that I'm on my way"
It was pretty innovative at the time (of the original iPhone technology demo which Apple accuhired).
Most of Siri's trouble is understanding speech (first) and command intent (second).
A chatbot interface doesn't help at all if the speech-to-text is busted.
An AI model for command/control speech-to-text would need actual training, e.g. the patterns which Siri understands today would need to be relearned within that model by example. To get back to what we have today. It is more useful for dictation, but would need tons of initial training and local training. How much would someone speak to their phone to teach it how to understand them?
The databases should not be that hard to update, but I do imagine that Apple does have hurdles for production changes, such as supporting the new feature in 20+ languages. The former-Apple-employee quote here seems a bit dated, since SiriKit allows third party developers to support incremental changes and the Siri recognition model is now able to be locally installed.