Computational workflows routinely execute tasks faster than a human scientist can understand and act on their outcomes, which means decisions about what tasks to run become outdated quickly. Artificial Intelligence (AI) algorithms have emerged as a route to adjust a workflow during operation, potentially increasing its effectiveness by learning which tasks may prove most informative. In this talk, we will discuss application patterns that integrate AI into scientific workflows and introduce software which simplify building such “AI steered applications.” The topics will include dissecting a materials design application built around a Generative AI model and middleware necessary to scale data-intensive AI workflows past thousands of GPU nodes.