Apart from XML::Twig you could also use XML::Rules like this:

use XML::Rules;

my $AC_n = 'CCC';

my $parser = XML::Rules->new(
  rules => [
    _default => 'as is',
     # by default keep both attributes and _content
    'Desc,What,AR' => 'content',
     # for those keep just the content
    '^AC' => sub {return ($_[1]->{n} eq $AC_n)},
     # only process the <AC> tags whose n attribute equals the $AC_n
    AC => '',
     # once processed, forget the contents of the <AC> tag
    AI => sub {print "Found AI: n=$_[1]->{n}, desc=$_[1]->{Desc}\n"; r
+eturn},
     # for each processed AI tag print the n attribute and Desc subtag
+.
     # thanks to the 2nd rule you don't have to write
     # $_[1]->{Desc}{_contents}
     # As this rule returns nothing, the contents of the tag are not r
+emembered.
  ],
);

$parser->parse($XML);
[download]

Unlike some of the suggested XML::Twig solutions this doesn't kep the whole document in memory, at each moment at most the data from a single <AI> tag and the attributes of the unclosed parent tags are kept.