The development of Artificial Intelligence (AI) solutions for preventive maintenance applications is a risky and resource-demanding process. Typically, there are several candidate solutions whose performance in transforming data into useful prognostic information is initially uncertain. These uncertainties can be managed by structuring the development process into multiple stages that help choose and implement the final solution. In this paper, we propose such a stage-gate process by using Robust Portfolio Modelling to screen increasingly detailed candidate solutions through four development stages and three decision gates. The development stages generate evidence on how the candidate solutions contribute to six development objectives that represent different financial and technical criteria. At the decision gates, decisions about the continuation/termination of candidate solutions are taken by identifying portfolios of non-dominated candidate solutions subject to time and budget constraints. Uncertainties are captured by admitting incomplete information about the criteria weights and scores of candidate solutions. We illustrate the process by considering the development of an AI solution for a train’s toilet door system. The process brings consistency to the development process and, among other benefits, helps mitigate the risk of missing the development objectives due to premature fixation on a single candidate solution. • A multi-stage decision process based on objectives and criteria is proposed. • Numerical scales are presented to help elicit comparable evaluation statements. • The process can be tailored to many kinds of AI-assisted prognostic tasks. • An illustrative example and a reflective discussion are presented. • The process offers benefits in terms of fostering organisational learning.

International audience