Same reason Turing machines are impractical; it's too low-level. But IMHO starting with "functional composition without side effects" (LC model) and adding the practical stuff gets you a lot farther than adding practical stuff to "everything is a side effect and functions don't exist" (the TM model).
Back in the days before parallel programming, large-scale software engineering, networks, cloud computation, etc it wasn't always clear that anything was wrong with the TM model. But given how computing works today it's clear (to me at least) that LC is a much, much better substrate for designing computation.
As an aside, 1958 was definitely back in those days. But that's when John McCarthy decided the TM model was lacking and said "Hey let's build a programming language on Lambda Calculus instead" and Lisp was born. And that's how John McCarthy was smarter than the rest of us.
Back in the days before parallel programming, large-scale software engineering, networks, cloud computation, etc it wasn't always clear that anything was wrong with the TM model. But given how computing works today it's clear (to me at least) that LC is a much, much better substrate for designing computation.