How do we make decisions, such as whether to order a red Bordeaux or a white Cabernet Sauvignon? The outcome of this decision will likely be one if you are thinking of combining it with oysters, and another if you are bound to combine it with boeuf bourguignon. This mundane example illustrates the flexibility with which we take context-dependent decisions. Context-dependent decision-making allows us to break free of rigid behavior and thereby to adaptively adjust our choices to different contexts. While it is notably impaired in major neuropsychiatric disorders such as schizophrenia, the neural underpinnings of this crucial cognitive ability are not yet comprehensively characterized. Recent research suggests that complex decision-making engages a wider network of regions across the brain, a view that we take in our research. With this aim, we combine machine learning with modern multiregional theories of decision-making and we approach this question in complementary ways: i) developed statistical methods to fit recurrent neural networks to invasive and large-scale recordings (from rats, monkeys or humans) acquired through collaborations or in-house and then reverse engineer the dynamical system extracted through gradient descent, ii) use our theoretical intuitions to directly build toy models that explain the neural data.