Principal types, which ML has, are useful because, for a given context, there is a most general type for any typeable term. This is a useful property to have for languages which attempt type inference.
Could you give some concrete example of a type inference engine and a typable term, but which does not have a principal type?