Classical model-atmosphere analyses of stellar spectra usually begin by measuring equivalent widths, and then proceed into a loop in which 1) model spectra are calculated for a set of abundances and atmospheric parameters, and 2) observed and computed spectra are compared and corrections to the abundances and parameters are inferred. Automated techniques have been developed to automate the measurement of equivalent widths, and some or all parts in the analysis loop. However, in order to tackle the massive datasets provided by the new spectroscopic surveys with dedicated telescopes, it is necessary to make some radical changes. It is argued that future analyses of stellar spectra should abandon the use of equivalent widths, and rely on tables of synthetic spectra that can be either interpolated extremely fast in minimum-distance optimization methods or used for training genetic algorithms. Examples of ongoing projects involving high-dispersion stellar spectra for a small sample and low-dispersion spectra for a sample of tens of thousands of stars are described.