Most AI projects fail. Some fail quietly before launch; some fail spectacularly publicly, becoming another media horror story about AI. Why does this happen? Because the current process for designing AI products and services is broken, especially when it comes to product strategy—what projects to pursue. But a new approach to designing AI is possible, one that instills more cooperation between designers, PMs, data scientists, and engineers.This talk walks through a new method that has been developed over many years at Carnegie Mellon University's Human-Computer Interaction Institute. This method uses elements of user-centered design and technology capabilities to find situations where moderate technical performance, high value, and low risk combine to make successful AI projects.This talk looks specifically about not just how, but also where and when AI should be deployed—and why most companies are doing AI wrong.