Three focal elements of knowledge-based system design are (i) acquiring information from an expert, (ii) representing the information in a system-usable form, and (iii) using the information to draw inferences about specific problem instances. In the artificial intelligence (AI) literature, the first element is referred to as knowledge acquisition, while the second and third are embodied in a system's knowledge base and inference engine, respectively. AI, however, is not alone in its concern for these issues. Researchers in several of the statistical decision sciences, notably decision analysis (DA), have also investigated them. This paper discusses the use of belief networks—a formalism that lies somewhere between AI and DA—as an overall framework for knowledge-based systems. Unlike previous work, which has concentrated on either the networks' mathematical properties or on their implementation as a specific system, this paper is oriented towards the concerns of general system design. Concrete examples are drawn from one medical system (Pathfinder) and from one financial system (ARCO1), and in particular, from a consideration of their similarities and differences. The design principles abstracted from these systems suggests a powerful, coherent design philosophy guided by the simple thought: form follows function.