Requirements in AI systems, like in traditional software systems, are rarely fixed at the start of a project, as changes are often necessary and difficult to avoid. However, AI software systems face unique challenges, including constant adaptation to evolving data patterns, uncertainty and ambiguity in requirements due to the involvement of multidisciplinary teams, and the frequent need for model retraining. These distinctive sources of requirement changes, combined with current development practices, further complicate the process. Current practices in AI system development often separate machine learning (ML) and non-ML module development, leading to a lack of systematic coordination during changes. To address the challenges in Requirement Change Management (RCM) arising from this lack of co-development, this paper proposes leveraging process modeling within a Model-Driven Engineering (MDE) approach to streamline RCM in AI development. Specifically, we introduce a Domain-Specific Modeling Language (DSL) tailored for the AI engineering process. This DSL aims to enhance traceability, standardize terminology across multidisciplinary teams, and provide the flexibility needed to manage the dynamic nature of requirement changes effectively.

International audience

Process Management Systems (PMS) offer effective means to coordinate tasks for various teams involved in complex projects. However, in practice, participants execute their tasks using applications within their workspace and manually report their activities to a separate PMS. This approach is not only time-consuming for end-users but also poses reliability challenges for the PMS in verifying the accuracy of reported task completion. This paper proposes an AI-augmented framework to establish seamless integration between PMSs and end-user workspace, which can automatically manage process execution by intelligently monitoring the tasks of process participants. We examined the utilization of our framework via a prototype pMage in a software implementation process. The results showcase the capability in managing process progress in various scenarios. The service effortlessly integrates the working environment with the corresponding process while maintaining its low-code applicability and independence from specific PMS or end-user workspace. This framework is anticipated to encourage end-users to embrace PMS as an integral part of their daily work, thereby unlocking the benefits of a process-aware workspace.

International audience