The inception of large language models has helped advance the state-of-the-art on numerous natural language tasks. This has also opened the door for the development of foundation models for other domains and data modalities (e.g., images and code). In this paper, we argue that business process data has unique characteristics that warrant the creation of a new class of foundation models to handle tasks like activity prediction, process optimization, and decision making. These models should also tackle the challenges of applying AI to business processes which include data scarcity, multi-modal representations, domain specific terminology, and privacy concerns. To support our claim, we show the effectiveness of few-shot learning and transfer learning in next activity prediction, crucial properties for the success of foundation models.