Membrane contact sites (MCSs) establish organelle interactomes in cells to enable communication and exchange of materials. Volume EM (vEM) is ideally suited for MCS analyses, but semantic segmentation of large vEM datasets remains challenging. Recent adoption of artificial intelligence (AI) for segmentation has greatly enhanced our analysis capabilities. However, we show that organelle boundaries, which are important for defining MCS, are the least confident predictions made by AI. We outline a segmentation strategy termed AI-directed voxel extraction (AIVE), which refines segmentation results and boundary predictions derived from any AI-based method by combining those results with electron signal values. We demonstrate the precision conferred by AIVE by applying it to the quantitative analysis of organelle interactomes from multiple FIB-SEM datasets. Through AIVE, we discover a previously unknown category of mitochondrial contact that we term the mitochondrial intrusion. We hypothesize that intrusions serve as anchors that stabilize MCS and promote organelle communication.