In this paper, we describe a computer architecture, which we call SecondOpinion, designed for automated, normative patient decision support over the World Wide Web. SecondOpinion custom tailors the discussion of therapy options for patients by eliciting their preferences for relevant health states via an interactive WWW interface and then integrating those results in a decision model. The SecondOpinion architecture uses a Finite State Machine representation to track the course of a patient's consultation and to choose the next action to take. The consultation has five distinct types of interactions: explanation of health states, assessment of preferences, detection and correction of errors in preference elicitations, and feedback on the implications of preference. A linear "summary model" speeds calculations of predictions from the decision model and makes it possible to dynamically calculate 95% confidence intervals for the marginal utility of each treatment option. Preferences for states are assessed in the order of their variance contribution to the models predictions in an iterative fashion. Only the states required to obtain a 95% Confidence Interval (CI) that excludes zero are assessed. In Monte Carlo simulation studies, the average number of utility assessments required for the 95% CI to exclude zero in an individual was 4.24 (SD = 1.97) out of 8 relevant health states. the SecondOpinion architecture provides an efficient, "discussion-like" experience leading to an individual-specific treatment recommendation. It may be a cost-effective approach to bring decision analytic advice to the bedside.