Current concepts conceive "breast cancer" as a complex disease that comprises several very different types of neoplasms. Nonetheless, breast cancer treatment has considerably improved through early diagnosis, adjuvant chemotherapy, and endocrine treatments. The limited prognostic power of classical classifiers determines considerable over-treatment of women who either do not benefit from, or do not at all need, chemotherapy. Several gene expression based molecular classifiers (signatures) have been developed for a more reliable prognostication. Gene expression profiling identifies profound differences in breast cancers, most probably as a consequence of different cellular origin and different driving mutations and can therefore distinguish the intrinsic propensity to metastasize. Existing signatures have been shown to be useful for treatment decisions, although they have been developed using relatively small sample numbers. Major improvements are expected from the use of large datasets, subtype specific signatures and from the re-introduction of functional information. We show that molecular signatures encounter clear limitations given by the intrinsic probabilistic nature of breast cancer metastasis. Already today, signatures are, however, useful for clinical decisions in specific cases, in particular if the personal inclination of the patient towards different treatment strategies is taken into account.