Today, most AI projects fail. Teams choose the wrong things to build. Based on 10 years of researching challenges in AI innovation as well as insights from teaching technical, business, and design students how to innovate with AI, we discuss how to brainstorm low-risk, high-value AI concepts, and things that are buildable and desirable. Neither user-centered nor technology-centered approaches work. We discuss how to integrate aspects of both while focusing on situations where moderate model performance generates value for users and service providers.