The separation of melt from crystals is a fundamental process driving the chemical differentiation of magmas and can lead to the formation of pockets of potentially eruptible magmas in highly crystallised magma reservoirs. While geochemical and geophysical evidence exists for the presence of such isolated pockets of eruptible melt, the processes that control their volume and spatial arrangement remain unclear. The Muroto sill in Japan provides an excellent opportunity to study these processes as it is perfectly exposed and shows clear evidence for melt segregation. We collected geochemical and structural data across the sill and performed thermal modelling to quantify extraction timescales and to constrain the range of crystallinity at which melt extraction occurred. Our data and calculations show that the middle–lower portion of the sill experienced melt extraction at crystal fractions between 0.65 and 0.8 over 100–150 years, until magma was too crystalline for further segregation to occur. We propose a new approach that can be used to invert measured geochemical profiles and identify the range of crystallinity at which melt extraction takes place. With this approach, the results we obtain for the Muroto sill can be generalised to magma reservoirs of different sizes and chemistries. Our calculations, and the comparison with natural magmatic systems, show that the volume of melt-rich pockets in a magma reservoir is proportional to the reservoir volume, while their spatial arrangement depends on the physiochemical properties of magmas. The results of this study increase our understanding of the factors controlling the distribution and volume of pockets of eruptible magmas in large magma reservoirs. Our calculations show that eruptible magma in dacitic and rhyolitic magma reservoirs, which are responsible for some of the largest eruptions on Earth, tend to be distributed in lenses of small volume within highly crystallised magma. Such architecture diminishes our capacity of identifying eruptible magma in large magma reservoirs such as Yellowstone using geophysical methods, and jeopardises our capacity of assessing the potential of a reservoir to feed a large eruption.