Agents that provide just-in-time access to relevant online material by observing user behavior in everyday applications have been the focus of much research, both in our lab, and elsewhere. These systems analyze information objects the user is manipulating in order to recommend additional information. Designers of such systems typically make the assumption that objects similar to the one being manipulated by the user will be useful to her. Our own experiments show that users do find many of the documents retrieved by a system of this type are relevant. Yet in the context of a specific task, users find fewer of these documents are useful. Our main point is that in order to make just-in-time information systems truly useful, we need to reexamine the "similarity assumption" inherent in many of these systems’ designs. In light of this, we propose techniques that bring modest amounts of task-specific knowledge to bear in order to perform lexical transformations on the queries these systems perform, thereby ensuring they retrieve not similar documents, but documents that are relevant and useful in purposeful and interesting ways.