Saaty's approach in the AHP framework divides inconsistent pairwise comparisons (PC) matrices into two categories, those with the acceptable inconsistency (with the consistency ratio equal to or under 0.10 threshold) and those with unacceptable inconsistency (above that threshold). The aim of this paper is to show that such a division is not appropriate, hence a new categorization of inconsistent matrices is proposed with respect to a satisfaction/violation of selected logical properties, such as the fundamental selection (FS) condition, the preservation of order preference (POP) condition, and the preservation of order of intensity of preference (POIP) condition. Moreover, a new non-linear optimization method for the derivation of weights (i.e. priority vector) is proposed such that the three aforementioned logical conditions are met. In the numerical part of the paper it is examined how frequently are the FS, POP and POIP conditions satisfied or violated for randomly generated PC matrices.