Old-style AI used rules and was deterministic, but was too human-intensive to deploy. What is the barrier now?

Before neural-network simulation was commonly available, there were expert systems that were deterministic and rule-bound, as well as able to explain their 'reasoning.' They were simply too expensive to create and update because you needed human experts and computer scientists to create them. Now we have AI that truly is at expert-level, but unreliable for a number of reasons. Why is no one pursuing either using the new AI to create expert systems, or at least using a much more hybrid approach?