Discovering the Right Things to Design with Artificial Intelligence

Advances in artificial intelligence (AI) enable impressive new technical capabilities: computers can diagnose diseases, translate between languages, and drive cars. Interestingly, today nearly 90% of AI initiatives fail; few projects survive until deployment. I argue that a lack of effective ideation leads teams to select suboptimal innovations to pursue. In addition, AI product teams fail to see low-hanging fruit, situations where simple predictive models can generate value for users and stakeholders. Currently, data science teams propose innovations customers do not want, while product teams ask for things AI cannot do. As AI capabilities become more pervasive and commoditized, discovering the right human problems to solve while mitigating potential harm remains a great challenge.

My research addresses this breakdown in early stage ideation and problem formulation. I studied practitioners and observed that teams better at ideating are more effective in developing AI solutions that generate value and minimize risk. Based on the industry best practices, I created new innovation processes and resources for helping cross-functional product teams effectively explore the AI solution space before selecting what to implement. I developed a taxonomy of AI capabilities and examples of these in product forms. These resources sensitize stakeholders to what AI can do and search for opportunities where these might be valuable. I developed a hybrid ideation method that blends technology-centered development and human-centered design. I conducted a preliminary assessmentof these resources and processes through case studies with innovation teams working in critical care,radiology, insurance, and accounting. Overall, this dissertation provides a glimpse into the future of human-centered AI innovation, where human needs and concerns are given equal importance as technical advances in deciding what to build with artificial intelligence.