Hamel Husain argues that if you can't evaluate your AI system, your users can't trust it either. He uses a data agent that answers business questions as an example: if you skip evaluation because it's too much work, you're hiding quality problems from the people relying on your tool.
I want to talk about one of the main reasons people say that they don't do evals. And that is, they say it just takes a long time. It takes forever to read a trace, to understand what's going on. And that's a lot of human hours that you just don't have time to do. For example, let's say you have an AI data agent that answers business questions. You want to know what happened to supply or revenue or answer any other business question that you would previously go to a data scientist for. This is like the most popular internal tool that I've seen so far. And the question is, how do you eval that? There's one important principle I think people need to keep in mind is if you can't eval your product, your product probably sucks. Why is that? It's because your users need to eval the product too. So if you present an answer from an AI to a user, let's say it's an answer to a business question, you need to figure out a way to convey trust.
