That sounds a bit reductionist, as if 'heavy use of monadic interfaces' would be kind of a silver bullet. There's a lot of concepts to internalize before being able to build large scale systems using categorical abstractions: Monads, monad transformers, free monads, applicatives, arrows, ...
When to choose what? Probably leave the choice to library-implementors mostly. Still, should the bar to implement a library be set so high? The middle ground seems to be what e.g. Rust is doing. No explicit categorical types, use functions where functions make sense, i.e. for most stuff that isn't I/O.