When making decisions, animals can employ many strategies, from relying on environmental stimuli to relying on internal biases. The roles of cortical regions in shaping these decision-making strategies remain unclear, often because we lack direct access to these internal variables. Here, we investigated how optogenetic inactivation of 29 motor, sensory, or association cortical regions affects decision-making strategies in 20 mice performing a virtual T-maze task in which they were rewarded for choosing the side with more visual cues. To determine mice’s underlying strategies on any trial, we fit a GLM-HMM, which determined which features (stimuli, side bias, trial history) predicted animals’ decisions. The model uncovered four primary states, including one in which decisions were primarily stimulus driven (‘engaged’), and two dominated by side bias (‘left’ or ‘right’). To study whether there was differential neural involvement across strategies, we determined how optogenetic inactivation impacted the decision-making features, separately during the time periods in which mice were in each GLM-HMM state. We found that, across all GLM-HMM states, optogenetic inactivation led to the stimulus being a smaller predictor of choice, corresponding to lower performance accuracy. The approach also revealed striking state-dependent effects that were unclear from averaging across all trials. Most notably, during biased states, inactivation of frontal regions (MOs and MOp) nearly eliminated side bias. Conversely, in the engaged state, frontal-cortex inactivation significantly increased bias. Interestingly, beyond impacting the animal’s current behavior, inactivation, especially of frontal regions, also significantly increased the probability of switching into the engaged state, suggesting a role of these areas in controlling transitions between behavioral strategies. These results 1) provide new insights into decision-making mechanisms, showing that cortical regions, particularly the frontal cortex, modulate bias and state-switching in a state-dependent manner, and 2) demonstrate a broadly applicable computational approach to understanding cortical involvement based on internal states.