It's about the hardest agentic problem that has measurable results and will be adopted (because it's an engineering culture) that I can think of. RunWare have some good insights on this - much of the problem is context layer improvisation and repetition, and the only way out is an awful lot of diagnostic development.
And anything less than 90% accuracy on causal analysis is more work than doing everything by hand.