Data are increasingly generated by many connected users, machines, and systems, which has led to explosive interest in collaborative analytics, such as federated learning, that borrow strength across diverse sources. To date, these efforts have focused mainly on predictive modeling. In this talk, I argue that collaboration, without untangling the nature of heterogeneity across users, can lead to swift and evident failures. Indeed, borrowing from the wrong sources can be worse than not collaborating at all. To this end, I present two frameworks that address two questions at the heart of collaboration: what to share, and whom to borrow from. First, we introduce Personalized PCA, a descriptive framework that extracts interpretable and identifiable features describing what is shared and what is unique across diverse datasets. Second, I bring language into the collaboration. When a new user has too little data to tell which collaborators are relevant, expert textual descriptions of the new system, often available yet overlooked, can fill the gap. Here, a large language model translates these descriptions into a prior over source relevance that guides learning when data are scarce and is gradually corrected as data accumulate. I conclude by highlighting future directions at the intersection of agentic AI, statistics, and decision making.