This is useful for both practical and teaching purposes: for practical because it keeps things simple in case the additional complexity isn't needed (e.g. scripting languages) and for teaching purposes because someone learns both ways (which are used in real world problems) while at the same time learning why one might be preferable to the other. And if you do the partial AST bit you even introduce the idea of an AST gradually by building off existing knowledge and experience the student has acquired.
And if you want to/need to later, you can trivially introduce an AST in those compilers by replacing the calls to the code-generator with calls to a tree builder, and then write a visitor-style driver for the code generation.
Instead of calling the code emitter, you call an AST builder.
Then you build a tree walker that call the code emitter.
At least the Wirth Oberon compiler was retrofitted with an AST by at least one student in Wirth's group as part of experiments with optimization passes.