The "starting point", or ancestral state, is chosen so as to obtain the most parsimonious tree.
Parsimony is used both for choosing the topology and for choosing the ancestral states for a given topology:
1) The parsimony score for a given topology is computed for the set of ancestral states minimizing the number of transformations.
2) Then, the topology minimizing this minimized number of transformations is chosen as the most parsimonious tree.