Hamel Husain argues that rating scales like 3.5 or 3.8 don't predict user satisfaction. Binary pass/fail evaluations are easier to align with human judgment, simpler to verify, and more reliable for measuring product quality.
Your AI eval says it's 95% helpful. However, your users still hate the product. This is because your metric is bad. Nobody knows what the difference between a 3.5 and a 3.8 score is. That's very ambiguous. It doesn't really mean anything, and AI is really bad at assigning scores, even if you give it a very detailed rubric. What you should do instead is make sure your evals have a binary outcome. Binary outcomes are easier to align with human beings, and they're easier to check, and they're also easier to generate labels for so that you can measure how good your eval is.
