With the launch of eROSITA (extended Roentgen Survey with an Imaging Telescope Array), successfully occurred on 2019 July 13, we are facing the challenge of computing reliable photometric redshifts for 3 million of active galactic nuclei (AGNs) over the entire sky, having available only patchy and inhomogeneous ancillary data. While we have a good understanding of the photo-z quality obtainable for AGN using spectral energy distribution (SED)-fitting technique, we tested the capability of machine learning (ML), usually reliable in computing photo-z for QSO in wide and shallow areas with rich spectroscopic samples. Using MLPQNA as example of ML, we computed photo-z for the X-ray-selected sources in Stripe 82X, using the publicly available photometric and spectroscopic catalogues. Stripe 82X is at least as deep as eROSITA will be and wide enough to include also rare and bright AGNs. In addition, the availability of ancillary data mimics what can be available in the whole sky. We found that when optical, and near- and mid-infrared data are available, ML and SED fitting perform comparably well in terms of overall accuracy, realistic redshift probability density functions, and fraction of outliers, although they are not the same for the two methods. The results could further improve if the photometry available is accurate and including morphological information. Assuming that we can gather sufficient spectroscopy to build a representative training sample, with the current photometry coverage we can obtain reliable photo-z for a large fraction of sources in the Southern hemisphere well before the spectroscopic follow-up, thus timely enabling the eROSITA science return. The photo-z catalogue is released here.