May be I'm misunderstanding you, but is this what you want?:

#! perl -slw
use strict;
use Data::Dump qw[ pp ];

my %categs;
push @{ $categs{ $_->[2] } }, $_->[1] while @{ $_ = [ split ' ', <DATA
+> ] };

pp \%categs;

__DATA__
31 John Consultation Charges 100 4/1/2012 10:39
805 Kennedy Consultation Charges 150 4/1/2012 11:17
805 Kennedy C-Reactive Protein 170 4/1/2012 11:56
805 Kennedy Complete Blood Count 150 4/1/2012 11:56
805 Kennedy Malarial 175 4/1/2012 11:56
805 Kennedy Mantoux Test 100 4/1/2012 11:56
805 Kennedy AZIBACT 1 MG SYP 28 4/1/2012 13:27
805 Kennedy FALCINILLE DRY SYP 105.15 4/1/2012 13:27
891 Ruth Consultation Charges 150 4/1/2012 12:05
891 Ruth C-Reactive Protein 170 4/1/2012 12:47
891 Ruth Complete Blood Count 150 4/1/2012 12:47
891 Ruth Mantoux Test 100 4/1/2012 12:47
891 Ruth X-Ray Chest 150 4/1/2012 12:47
[download]

Producing:

C:\test>junk59
{
  AZIBACT      => ["Kennedy"],
  "C-Reactive" => ["Kennedy", "Ruth"],
  Complete     => ["Kennedy", "Ruth"],
  Consultation => ["John", "Kennedy", "Ruth"],
  FALCINILLE   => ["Kennedy"],
  Malarial     => ["Kennedy"],
  Mantoux      => ["Kennedy", "Ruth"],
  "X-Ray"      => ["Ruth"],
}
[download]

Do you have some kind of "dictionary" which tells you into which category or categories each type of service belongs? If so, then you just have to read each service and check it against the dictionary to find out into which category or categories each service belongs. Once you have done that, you check the number and type of categories for each client and put that info in some kind of "scoring" formula to find the most appropriate category.

If you do not have a "services-to-categories" dictionary then things become much more difficult and I really do not have a good and simple solution. I once applied Bayesian statistics on a similar problem (though only with a few broad categories to put the records in) and it worked "somewhat". I got about 80% correct categorizations (and thus 20% totally wrong), but it was enough for my purpose. If I trained the algorithm a bit more I might have gotten better results. Modules such as Algorithm::NaiveBayes or AI::Categorizer::Learner::NaiveBayes are worth taking a look at.