As machine learning architectures stabilize, there is a methodological shift towards data-centric AI (DCAI)-designing training data characteristics while keeping the model constant to achieve desired behavior and performance. We argue that this shift is a promising path forward to realizing human-centered AI. Based on qualitative inquiry (interviews and in-lab co-design studies) with industry practitioners, we find that data is under-emphasized in current AI development practices and is optimized for engineering tasks rather than end-users. Insights from our studies show that HCI practitioners leverage well-established user research and design techniques to anchor AI development around human needs. End-user data in the form of design probes serve as the lingua-franca for HCI and AI practitioners to collaborate on system design.