This paper describes three versions of an expert system for marine liability insurance underwriting. The first version, built on a PC, was the result of a series of knowledge engineering sessions with an underwriter of an insurance company. The prototype was accurate, but it asked too many marginally relevant questions, without regard to the time and business constraints of the user. The second version was developed to address this concern. It initially asked only a subset of the relevant questions to arrive at a premium. If the premium was too high, the prototype attempted to reduce it by performing a sensitivity analysis on the results of the first round. This analysis was driven by meta-rules. The third version was built on a lisp machine using a graphical expert system-building tool. The graphical approach and greater flexibility of the tool provided the user with an overview of the domain and permitted him to have more control over the data. Any implementation of an expert system for marine liability insurance underwriting should incorporate the lessons learned from each of the three versions of the prototype. In addition, sensitivity analysis and a graphical interface could be useful for expert system domains other than insurance underwriting.