The ubiquitousness of data and the emergence of data-driven machine learning approaches provide new means of creating insights. However, coping with the great volume, velocity, and variety of data requires improved data analysis methods. This dissertation contributes a nascent design theory, named the Division-of-Labor framework, for developing complex machine learning systems that can not only address the challenges of big data but also leverage their characteristics to perform more sophisticated analyses. I evaluate the proposed design principles in three practical settings, in which I apply the principles to design machine learning systems that (i) support treatment decision making for cancer patients, (ii) provide consumers with recommendations