AI has significantly enhanced the abilities of knowledge workers in various tasks, e.g., composing emails, summarizing documents and videos, and data analysis. Despite AI systems reducing much of the workload, there are still challenges for human users to interpret and incorporate AI-generated results into their work. We identified two primary types of challenges, framed using the classical human-computer interaction (HCI) notion of “gulf of evaluation”: (1) the inherent inaccuracies of AI models; (2) the lack of explanations for AI-generated output. To bridge these gulfs, we need to develop novel human-centered AI systems and interfaces. This dissertation proposal explores the development of systems and techniques to bridge the gulf of evaluation in human-AI interaction. We focus on two types of knowledge workers: content creators and data analysts, in four problem domains: mixed-media tutorial creation, workspace text-based communication, event sequence data analytics, and machine learning model validation. We share specific findings from empirical investigations with real users, and distill generalizable lessons and strategies on designing human-AI interaction to bridge the gulf of evaluation.