One last addendum on the accuracy topic.
Here’s a sample which should illustrate how incredibly cumbersome editing may get using Subdivision Surfaces, as soon as one wants to do precision work. The reason in short is that in Subdivision Modelling one generally can not cut away or add geometry wherever one likes. Never.
Either one was smart enough to lay out the cage appropriately in the first place, or one needs to rebuild the topology to accomodate the change. This little demo clip (from a series of generally very recommendable series of introductory videos to mesh editing and its topology demands) demonstrates the basic concept of combining several input shapes to one. Here’s the entire playlist.
Say this egg shape was some plastic part we were designing and one needed to add some dome structure for mounting several parts together. The goal is to to maintain a valid SubD-Topology with unaltered curvature, boolean operations, triangles and Ngons which destroy loop based topology were not allowed.
If you want to give this a try – go ahead: Use Modo or Catia, Tsplines, interpolated SubD or mesh approximated subdivision. Redo everything from scratch if that makes sense – I’ll buy anyone a 
who comes up with a good looking solution and needs less than hours to do this. Here’s a Rhino 5 file with the cages. sample_modelSubD.zip
I personally would never even try doing such by hand – here’s at least a quick cheat which roughly demonstrates how a proper mesh reconstruction had to look.
Even if one was such an accomplished SubD modeller – what had one actually done? The input shapes were super simple, low poly and perfectly editable, ideal for freely playing around and finding shapes. The outcome is greatly more complex and determined to stay that way, quite the way a dense Nurbs surface doesn’t leave much room for further experimentation.
How much nicer would things be for the Designer if one could feed those individual volumes into Grasshopper (possibly even Nurbs geometry and meshes combined) and GH spit out the end result as a mesh which was usable for rapid prototyping, the creamtopping obviously was intersection manipulation á la MeshFusion – nudge nudge @DavidRutten As soon as one figures that one is getting somewhere there was still time to build an accurate model. Using Nurbs.