We have investigated in detail the uncertainties in deriving beryllium abundances for metal-poor stars from the BeII resonance doublet at λ 3131A. These uncertainties are related to the quality of the observations, the physical data used, and the adopted stellar parameters. We have taken into account the NLTE effects on the lines formation for the Sun and a sample of six stars with metallicities in the range -0.4>=[Fe/H]>=-2.7. NLTE and LTE calculations give nearly the same equivalent widths for the Sun, confirming the previous results of Chmielewski et al. (1975). The extension of the NLTE analysis to metal-poor stars indicates that the absolute NLTE corrections to the beryllium abundances are lower than 0.1dex. The six selected metal-poor stars have been previously observed by other authors, and show a dispersion of ~0.2-0.5dex in their Be abundances (which can be as large as ~1dex when including the error bars). Our analysis of these stars provides consistent abundances with uncertainties in the range 0.2-0.3dex, which are basically related to the errors in the adopted stellar parameters (mainly surface gravity). These uncertainties and NLTE corrections <~0.1dex should be taken into account when using LTE beryllium abundances for primordial nucleosynthesis and Galactic chemical evolution studies.