We have recently developed a speech-recognition system for automation of telephone-service orders, and tested the system through four months of regular use by GTE customers. In this paper, we use quantitative data (response content) to assess the effectiveness of our structured transaction model, and evaluate the extent to which natural queries yielded responses in the predicted linguistic form. Our results showed the structured model to be a successful approach to transaction automation. The results also show that customer responses to our designed queries were both predictable in form and quite limited in variety.