Fast and accurate decisions are fundamental for adaptive behaviour. Theories of decision making posit that evidence in favour of different choices is gradually accumulated until a critical value is reached. It remains unclear, however, which aspects of the neural code get updated during evidence accumulation. Here we investigated whether evidence accumulation relies on a gradual increase in the precision of neural representations of sensory input. Healthy human volunteers discriminated global motion direction over a patch of moving dots, and their brain activity was recorded using electroencephalography. Time-resolved neural uncertainty was estimated using multivariate feature-specific analyses of brain activity. Behavioural measures were modelled using iterative Bayesian inference either on its own (i.e., the full model), or by swapping free model parameters with neural uncertainty estimates derived from brain recordings. The neurally-restricted model was further refitted using randomly shuffled neural uncertainty. The full model and the unshuffled neural model yielded very good and comparable fits to the data, while the shuffled neural model yielded worse fits. Taken together, the findings reveal that the brain relies on reducing neural uncertainty to regulate decision making. They also provide neurobiological support for Bayesian inference as a fundamental computational mechanism in support of decision making.