Several pieces of evidence have been pilling up in the literature leading towards an evolutionary scenario for AGN, moving past the standard unified model. Part of this amounting evidence is the little overlap of AGN samples selected with various identification criteria. Given the rarity of the AGN population, large and complete samples are needed to assess such an evolutionary model and clarify the growth of black holes and a potential coevolution with their host galaxies.
X-rays are undoubtedly an optimal AGN identification method as revealed by XMM-Netwon, Chandra, and now eROSITA. However, deep all sky surveys such as LSST and Euclid will provide at least one order of magnitude larger galaxy samples compared to the most optimistic predictions of eROSITA. I will present three recent approaches using multiwavelength datasets tailored to identify AGN with machine-learning methods. These methods have been developed and tested on available optical, near infrared and X-ray public datasets and will reach their true potential with the combined power of LSST, Euclid and eROSITA.