Corporate diversification theory is being reshaped by data and AI, yet evidence on how multi-business firms should organize AI across their portfolios remains scarce. We address a corporate-strategy question that is increasingly central in the digital era: should AI be implemented as a corporate-wide policy or tailored at the business-unit level, and does the answer depend on the nature of the product offering? Drawing on the theory of the firm and recent arguments about digital diversification, we theorize that the performance impact of AI implementation locus is contingent on whether business units manufacture smart products—offerings with embedded narrow AI that operate within interconnected value ecosystems. Using a unique dataset of 667 manufacturing product lines from Spanish firms, we show that across the full sample neither implementation locus dominates unconditionally. However, the interaction with smart products reveals a clear pattern: for non-smart product lines, corporate-wide AI implementation is associated with superior market performance, consistent with internal economies of scope; for smart product lines, business-unit AI implementation yields superior performance, consistent with ecosystem-facing specialization and interoperability demands. The findings contribute to the emerging literature on digital diversification by showing that AI governance should be treated as a portfolio design problem, centralizing what truly scales while decentralizing what must remain close to product architectures, customers, and ecosystem interfaces.