Shafer's theory of evidential reasoning has recently received much attention as a possible model for probabilistic reasoning in expert system applications. This paper discusses the particular difficulties of implementing Shafer's belief functions in the context of the most common form of expert system, rule-based systems. The two most important problems are: the representation of the expert's subjective degrees of belief corresponding to his expressed rules, and the computational complexity of the inference mechanism for combining evidence. We argue that a potential approach for dealing with both problems is given by introducing constraints on the structure of the belief functions. These constraints, along with the expressed rules and the elicited belief values, form the expert's total knowledge.