When learning something for the first time, working memory is needed to keep in mind the reasons and rules behind our actions and to evaluate how successful our strategy is. After discovery of the correct rule, the working memory load is reduced, and with extended training, automatic responding develops, which skips this deliberative process. But even as we learn a new procedural skill without relying on conscious deliberation, only extended training can lead to quick and efficient automatic responses. An intriguing possibility is that the qualitatively and neurobiologically distinct declarative and procedural learning systems are ultimately training up the same automaticity system. We tested this idea with fMRI studies. Initial rule-based task performance was correlated with activation in PFC, the hippocampus, and the head of the caudate nucleus, whereas early procedural skill training depended heavily on the putamen, where synaptic plasticity is reward dependent. By session 20 however, activation in all of these areas no longer correlated with performance. Instead, only cortical activation (e.g., in premotor cortex) was positively correlated with response accuracy in both tasks. This led us to theorize that both category learning systems have similar goals—namely, to train automatic cortical-cortical representations. These cortical projections from the triggered sensory areas go directly to the premotor areas that initiate the behavior. This efficiency is akin to traveling on speedy superhighways, subverting the need to stop at neural regions specialized in deliberative or reward-based learning.