We introduce SAM3D, a new approach to semi-automatic zero-shot segmentation of 3D images building on the existing Segment Anything Model. We achieve fast and accurate segmentations in 3D images with a four-step strategy comprising: volume slicing along non-orthogonal axes, efficient prompting in 3D, slice-wise inference using the pretrained SAM, and recoposition and refinement in 3D. We evaluated SAM3D performance qualitatively on an array of imaging modalities and anatomical structures and quantify performance for specific organs in body CT and tumors in brain MRI. By enabling users to create 3D segmentations of unseen data quickly and with dramatically reduced manual input, these methods have the potential to aid surgical planning and education, diagnostic imaging, and scientific research.