Breast cancer specific mortality results from tumour cell dissemination and metastatic colonisation. Identification of the cells and processes responsible for metastasis will enable better prevention and control of metastatic disease, thus reducing relapse and mortality. To better understand these processes, we prospectively collected 307 patient-derived breast cancer samples (n = 195 early breast cancers (EBC) and n = 112 metastatic samples (MBC)). We assessed colony-forming activity in vitro by growing isolated cells in both primary (formation) and secondary (self-renewal) mammosphere culture, and tumour initiating activity in vivo through subcutaneous transplantation of fragments or cells into mice. Metastatic samples formed primary mammosphere colonies significantly more frequently than early breast cancers and had significantly higher primary mammosphere colony formation efficiency (0.9 % vs. 0.6 %; p < 0.0001). Tumour initiation in vivo was significantly higher in metastatic than early breast cancer samples (63 % vs. 38 %, p = 0.04). Of 144 breast cancer samples implanted in vivo, we established 20 stable patient-derived xenograft (PDX) models at passage 2 or greater. Lung metastases were detected in mice from 14 PDX models. Mammosphere colony formation in vitro significantly correlated with the ability of a tumour to metastasise to the lungs in vivo (p = 0.05), but not with subcutaneous tumour initiation. In summary, the breast cancer stem cell activities of colony formation and tumour initiation are increased in metastatic compared to early samples, and predict metastasis in vivo. These results suggest that breast stem cell activity will predict for poor outcome tumours, and therapy targeting this activity will improve outcomes for patients with metastatic disease.