If you have good tests, certain types of change can be merged without manual testing. One problem specific to AI is that it has a tendency to game/bypass/nerf/disable tests, as opposed to actually making the code do the correct thing.