I think I said that wrong; the activation function of neural networks doesn’t multiply the weighted sum by the sigmoid function; it uses the sigmoid function to map the weighted sum onto an interval.
But the main point remains, which is that it is nonlinear.