I think I found my answer.
Assuming the input program is not fully CPS, at least one procedure return point will have to be converted into a continuation by the CPS conversion. So this continuation is both introduced by the conversion and necessary. Because it is necessary, you would always need to do this, also when converting by hand for example. Therefore, administrative redexes are only those lambdas introduced by the CPS conversion that aren't really necessary (my second definition).
I found a paper that explains it like this (emphasis mine):
The naîve λ-encoding into CPS,
however, generates a quite impressive
inflation of lambdas, most of which
form administrative redexes that can
be safely reduced. Administrative
reductions yield CPS terms
corresponding to what one could write
by hand. It has therefore become a
challenge to eliminate as many
administrative redexes as possible, at
CPS-transformation time.
Still, any remarks or suggestions welcome of course.