Automated vehicles can increase network capacity by leveraging connectivity with other vehicles and the infrastructure, but come at a larger operating and maintenance cost than other human-driven vehicles. This study investigates the trade-off in cost and capacity effects of automated vehicles when a social planner decides the optimal fleet mix for serving a given network demand. We present a system optimal assignment problem that is a non-linear program with linear constraints. Using second-order analysis, we label each link of the network as either “convex” or “concave”. The classification is done prior to any traffic assignment and is independent of demand patterns. According to the classification, a social planner benefits from increasing (decreasing) the automated vehicle ratio of traffic in convex (concave) links. We show that the proposed system optimal problem is non-convex and develop a Benders decomposition algorithm with a master problem that decides path flows and a sub-problem that finds the optimal automated vehicle ratio of each path. We further discuss the system optimal properties in three stylized networks with parallel, series, and mixed topology, and a larger Sioux Falls network. Results show that in parallel networks, the convex (concave) links have an automated vehicle ratio of one (zero), whereas in series networks, the optimal automated vehicle ratios depend on several network parameters.