Why experimentation—and Excel—might hold the k