The main issue with expert systems appeared to be not the amount of necessary rules but pathway dependency on representation model for the problem domain. If it

The main issue with expert systems appeared to be not the amount of necessary rules but pathway dependency on representation model for the problem domain. If it mapped well, it could perform exceptionally well. If not, you got into all sort of combinatory troubles real quick and had to wing it (or as they used to say in the 1980s, to use heuristic rules). The combinatory bit due to its branching nature lends itself poorly to leveraging SIMD hardware as statistical solutions do. And getting a right representation model is a substantial challenge. Everyone knew the importance of that ('knowledge engineering') by the onset of AI Winter but nobody had any good methods.