Abstract Humans can flexibly change rules to categorize sensory stimuli, but their performance degrades immediately after a task switch. This switch cost is believed to reflect a limitation in cognitive control, although the bottlenecks responsible for this remain controversial. Here, we show that humans exhibit a brief reduction in the efficiency of converting sensory inputs into decision evidence immediately after changing rules in perceptual tasks. Participants performed a flexible face categorization task in which they classified parametrically generated face stimuli based on one of two rules, switching every few trials. Although participants were always informed of a rule switch by a context cue, they showed a specific pattern of increase in reaction times, regardless of the time they were given to prepare for the switch. Psychophysical reverse correlation and computational modeling revealed a reduction in sensory weighting immediately after a rule switch, which recovered within a few hundred milliseconds after stimulus presentation. Furthermore, this cost depends on the sensory features being switched, suggesting a major bottleneck at the stage of adjusting the sensory information flow. We propose that decision-making circuits cannot fully adjust their sensory readout based on an abstract context cue alone, but rather require the presence of an actual stimulus to tune it, leading to a limitation in flexible perceptual decision making.