They’re useful when you need atomic changes across multiple documents/collections. The cost is more overhead and reduced performance compared to single-document operations; you should model data to avoid needing multi-doc transactions when possible.