More and more, I'm thinking about this idea to provide a bounding box that encapsulate all the animations is not as bad as said ...
Please proove me I'm wrong (with pleasure )
As Buckshag mentions in the following...
The method using keyframes would only work if you do not blend or combine motions or do not do any procedural things.
...If you use blending or additive animations (which you most likely will), then only taking the bounds of each pure keyframe and collecting only the bigger dimensions cannot account for all possible scenarios. The addition of two extreme frames may result in something that pierces the pre-calculated bounds.
However, the odds of this making the entity pop in or out of view in the middle of the game are very low, especially because you can increase the bounding box by a fixed amount and be conservative.
Unity seems to use the simple keyframe calculation at import: http://docs.unity3d.com/Documentation/Components/class-SkinnedMeshRenderer.html. Note their explanations on how additive animations can alter the result.
You can also consider using a bounding sphere instead, which is a more conservative volume.