Fully agree. I found for most of my work with LLMs and Finance, ~90% of the time a high or low confidence score was accurate. There's the occasional ambiguous case, and that'll happen, but the engineering work that comes with building a classifier makes it not practical for my usecases.