In recent years, digital twinning has become a mainstay of additive manufacturing (AM) processes, supporting process stability and the repeatability of the manufactured parts. Developments in digital twinning have also gone hand-in-hand with improvements in the manufacturing technologies themselves, leading to increasingly controlled processes, particularly in the case of smaller scale metal or polymer AM processes. In contrast, concrete AM is typically a less predictable process with many variables, and often involving manual intervention, e.g., for inserting reinforcements. Additionally, the physical surroundings are often more complex due to the presence of workers and other objects, as well as significant variations in the environmental conditions such as ambient temperature, humidity, and light, all of which influence the material properties of the concrete. Nevertheless, there remains a need for automated digital twinning of the process to ensure the quality of the manufactured objects. This demands the development of flexible and adaptive tools that leverage intelligence to deal with both expected and unexpected situations. In order to address these challenges, we propose a framework for building a digital twin based on a number of data inputs including camera images and environmental sensors, as well as nominal geometries and tool paths. The approach is centered on using Bayesian Networks to combine directly observed variables with inferences from foundational AI models for image segmentation and vision-language models. The result is an explainable model of the concrete AM process, unpacking its variance into factors that are grounded in real-time data and expert knowledge.