Astrometric binaries are both a gold mine and a nightmare. They are a gold mine because they are sometimes the unique source of orbital inclination for spectroscopic binaries, thus making it possible for astrophysicists to get some clues about the mass of the often invisible secondary. However, this is an ideal situation in the sense that one benefits from the additional knowledge that it is a binary for which some orbital parameters are somehow secured (e.g. the orbital period). On the other hand, binaries are a nightmare, especially when their binary nature is not established yet. Indeed, in such cases, depending on the time interval covered by the observations compared to the orbital period, either the parallax or the proper motion can be severely biased if the successive positions of the binary are modelled assuming it is a single star. With large survey campaigns sometimes monitoring some stars for the first time ever, it is therefore crucial to design robust reduction pipelines in which such troublesome objects are quickly identified and either removed or processed accordingly. Finally, even if an object is known not to be a single star, the binary model might turn out not to be the most appropriate for describing the observations. These different situations will be covered.