I think evaluation is often the real bottleneck in applied AI, especially for teams shipping RAG systems or agents. A stronger model can be swapped in quickly; establishing whether answers are correct and catching subtle quality declines takes substantial work.
My view is that Evaluation, rather than the model, is rarely the bottleneck in applied AI.
I raised the question of how people shipping RAG systems or agents evaluate them, mentioning human review, LLM-as-judge, or another approach. I did not state which approach I use.
I shared this opinion in an X post dated 2026-11-05 19:13 UTC. The post reported no evaluation results or quantitative measurements.