A carefully designed stratified sampling scheme is proposed for detecting the maximum possible errors in highly skewed accounting populations. The further possibilities concerning the use of this scheme under optimum allocation are explained. The problem of the choice of the stratum boundaries is discussed and the optimal stratification cut-off points are obtained by minimizing the Var (?) under the assumption of exponential distribution. The approximate methods as well as an exact method for the general case of 'k' strata are presented. A table showing the connection between 1+1' strata values is also given.