Teaching to learn and learning to teach: using machine training to develop artificial intelligence partners in new product development

Artificial intelligence (AI) is increasingly shaping how innovation is developed, and new product development (NPD) is an important domain at the centre of this shift, given its data-rich, high-uncertainty contexts. In this paper, we focus on the design and training of AI systems in which human expertise is encoded into machine models and the future scope of human and machine agency is determined. Drawing on organizational learning theory, we argue that training strategies defining who provides labels, which data are used, and how feedback is generated and reused, constitute design decisions that differ in how feedback is recursively structured, where knowledge is located, and how deeply expert judgment is codified. We develop and test three strategies, experiential, nudging, and generative, and instantiate them through computational simulations in the domain of geotechnical services for underwater exploration. We further assess the cross-context portability of the most effective strategy by redeploying the trained model in a second NPD setting. Our results show that strategies that recursively incorporate expert feedback across training cycles produce reliable, transferable models, whereas those that decouple human expertise from the training process converge prematurely on narrow solution spaces. These findings suggest that how firms structure the training of AI systems directly shapes model reliability and transferability, and provide a basis for theorizing how organizational routines, competencies, and relational forms of AI agency may develop as AI systems become embedded in NPD processes.