Most e-commerce sites today are little more than electronic catalogs of product offerings. Consumer input is limited to requirements questionnaires, search engines, and accepting or rejecting particular offerings. But in complex purchases, such as real estate, cars, or computers, it is often difficult to specify exactly what you want, and priorities and preferences often change in the process of exploration. We are investigating software agents that use machine learning, context sensitivity, and predictive interfaces. We would like these agents to act as an advisor to a consumer much as a real estate agent or travel agent would, implicitly inferring general preferences from a history of relatively unconstrained reactions to specific examples.