Successful deployment of speech understanding systems demands an understanding of all relevant aspects of human/machine interaction. By determining the factors which most significantly affect user behavior, and quantitatively describing the effects of varying these factors, we can more accurately control and predict a system's ultimate performance under real user conditions. This paper describes the design of an automated system, Money Talks, which will determine speech recognition and application requirements in a real user, service-providing environment: a telephone-based information service. We discuss the reasoning and implications underlying our design decisions to investigate certain aspects of interaction, while not investigating others. We also discuss the trade-offs between "naturalness," efficiency, and the effect on speaker compliance associated with wording variations in prompts. Lastly, issues related to the design of such trials, such as the need for barge-in technology, realistic simulations of graceful failure, and operator-handoff procedures, are also addressed.