CAD files often have quite a bag of problems if visualization is needed. They are all created with different targets. Based on their core modeling technique, most of the higher-end CAD systems run mathematically based. From that data, polygon-based models are retrieved. Here the settings are the key to getting either bad or better models.
What I am trying to say, there is a wide variety of CAD systems and an even wider variety of polygon results. This makes it so hard to say anything in general.
My typical advice is to get in contact with the person who creates the CAD data in the first place and try to make him/her understand what you need. As mentioned, what is done in CAD doesn't mean targeting visualization in the first place. So, don't fix what could be created from the start in a better way. For a CAD operator, that is often the same work either way. I'm sure everyone likes to get a thank you instead of perhaps a cool OK. So, I stick with the idea that the workflow has room for improvement when I get a question like yours.
I was an office principal/manager/lead designer in the early '90s, and my first project was to connect my teams of architects with our CAD department. That was quite new back then. It took a while; I trained the CAD people a little bit in Architecture, and the architects went with me to Nemetschek and enjoyed some CAD training. Building bridges. After that, the CAD room was overbooked, and everyone was happy. Since then, I have seen many times that things have options to get better. Try it.
Yes, my roots with Nemetschek (Allplan, ArchiCAD, and Vectorworks) are deep, and I love them. It was fun to see Maxon joining this great company. In short, I believe there is always something possible.
There are methods to improve, but again, I can't say anything in general. I need to see the model. Sometimes it is as easy as using the Untriangulate option or combining tools. This is kind of experience based. Take a copy and mess around with it. Of course, ask, but please with an example.
I'm happy to look into it. After a while, you will see what is possible. As everything has quality, it needs some time to develop.
The key problem would be that the Polygon where the Root is will be lifted. Hence the results look somewhat chaotic.
If you could connect the points of the hair instances with the "main body", then the Polygon would be inside of it. Meaning you have three polygons sharing a single edge. This will not provide excellent results.
There is, of course, the idea of having two "main bodies" animated in the same way, preferably via PLA.
With this, one would have the Root information, and the other would receive the results. However, the Hair instances are generated, and I'm unclear on how to combine them and have those working with SDS.
A good mesh is based on knowledge and/or manual work.
It is not often that a roof gets a deformer treatment. With lots of polygons, deformers might work; the kind of deformation indeed dictates the structure of the mesh. Just lots of polygons are often slowing down the processing. So, we end up doing it manually anyway.