When it is not going to view in detail and the geometry of vertices isn't important.
As you said, there is no extra vertices created, which mean, if there isn't a geometry, rendering engine can't render on that pixel. Commonly seen in early game when we can't afford high vertices computing.
As a same result, AO, shadow, light bounces cannot be accurately calculate due to the rasterized geometry (Although there is some trick that can compute them in additional process).
So how do we decide which one to use:
- Is our scene need to be realistic as possible as we can?
- Can the rendering engine/hardware afford such high vertices? Or high resolution normal map?
- Can we obtain the normal map?
- Do we have time to bake normal map? (especially you don't want to waste time on tweaking your map and vertices. eg. making animation)
- Is the rendering time important? (Realtime Game, VR, AR, news animation)
There is much more thing that we should consider about, including what pipeline are the artist using, which might be the most important factor in current decision. And I don't think an short Q&A post can cover all the issue in this problem.