Abstract: Aspects of the disclosure relate to training and using machine-learning models to determine an optimal number of robotic process automation (RPA) bot instances to process a work queue. In some embodiments, a computing platform train a machine learning model based on RPA using process step information and historical execution log data. Thereafter, the computing platform may receive a current work queue of items to be processed, and determine, using the machine learning model, an optimal number of RPA bot instances to process a current work queue. The computing platform may send the determined optimal number of RPA bot instances to a RPA bot orchestrator on a virtual bot host server, causing the orchestrator to instantiate at least one bot instance corresponding to the determined optimal number of RPA bot instances to process items from the current work queue using the at least one bot instance.