One method that could possibly work is to use the clouds you have to
estimate the pdf of C-points and A-points, then use something like the
likelyhood ratio derived from these pdf's as indicative of which class a
new point is likely to belong to.
One way of estimating the pdf could be to use Kernel Density Estimation
(routines to do this are included with the Matlab distribution I believe
and if not included are freely available).