I like defining AI as a catch-all for higher-order solutions. Rather than defining a specific process for taking in input and producing the desired output, you define a process for taking in input to produce a process that takes in input and produces the desired output. That ends up including a lot of boring applications, like SAT solvers, Bayesian statistics engines, as well as the more hip deep learning stuff.

ML is the specific case where the inputs to both the higher level and base level processes are similar, and the goal is for the application to identify patterns to apply to specific cases.