We propose a new method for input delay compensation of nonlinear systems that can be linearized using restricted feedback. Our approach is based on an inversion procedure that finds the input that tracks asymptotically the desired output. This kind of inversion schemes, typically, needs an estimate of the future of the system's state. In order to compute this estimate, instead of integrating the original system (which might be nonlinear or unstable), we integrate the desired error dynamics (which is both linear and stable, at least asymptotically). Our approach is validated numerically on an academic nonlinear pendulum example.