The Fermi Gamma-ray Space Telescope is producing the most detailed inventory of the gamma-ray sky to date. Despite tremendous achievements approximately 25% of all Fermi extragalactic sources in the Second Fermi LAT Catalogue (2FGL) are listed as active galactic nuclei (AGN) of uncertain type. Typically, these are suspected blazar candidates without a conclusive optical spectrum or lacking spectroscopic observations. Here, we explore the use of machine-learning algorithms - Random Forests and Support Vector Machines - to predict specific AGN subclass based on observed gamma-ray spectral properties. After training and testing on identified/associated AGN from the 2FGL we find that 235 out of 269 AGN of uncertain type have properties compatible with gamma-ray BL Lacs and flat-spectrum radio quasars with accuracy rates of 85%. Additionally, direct comparison of our results with class predictions made after following the infrared colour-colour space of Massaro et al. (2012) show that the agreement rate is over four-fifths for 54 overlapping sources, providing independent cross validation. These results can help tailor follow-up spectroscopic programs and inform future pointed surveys with ground-based Cherenkov telescopes.