The Loewner framework for model reduction is extended to the class of bilinear systems. The main advantage of this framework over existing ones is that the Loewner pencil introduces a trade-off between accuracy and complexity. Furthermore, through this framework, one can derive state-space models directly from input-output data without requiring initial system matrices. The recently introduced methodology of Volterra series interpolation is also addressed. Several numerical experiments illustrate the main features of this approach.