The hypothesis of this paper is that AI models must be trained on data generated by workers (both in- and out-of-house), and that this complementarity between labor and data fundamentally alters AI’s effects on the labor market. We develop this hypothesis in three steps. We link the Annual Business Survey (ABS) and Business Trends and Outlook Survey (BTOS) to individual tax records and document (1) the importance of in-house R&D for AI adoption, (2) the importance of data and talent for AI adoption, and (3) the heterogeneous effects of AI adoption on wage outcomes. We find that exposed workers gain relative to non-exposed workers on average, particularly so in occupations where firms conduct significant in-house R&D on AI (i.e., where AI adoption is not ‘out-of-the-box’). Second, we build a theory of AI adoption as an experience good (i.e., AI combines external and internal data on worker tasks) and we integrate this learning process into a directed search model with skill-weights. We allow firms to adapt their skill-weights, endogenously giving rise to new work, and we allow workers to retrain. We then estimate the model and compute labor transition paths as AI improves over the next 20 years. A 90% reduction in adoption costs, a 9-fold increase in exposure, and a 90% reduction in data obsolescence fail to generate widespread displacement. Workers retrain and remain complementary with data, bounding unemployment below 6.5% in the long-run. Lastly, an AI insurance fund for fully automated workers facilitates retraining and yields significant welfare gains.