> they are almost never used in any real automatic differentiation system

They're efficient enough for first-order derivatives. For example they are used in Ceres, Google's library for non-linear least-squares optimization

https://ceres-solver.googlesource.com/ceres-solver/+/master/...