TRACE: LLMs search and retrieve rules, but still don't apply them
A benchmark for whether AI can execute complex, multi-step enterprise processes correctly, from start to finish. We tested the whole field, and the best system gets only 38% of tasks fully right.