Jakob Nielsen has written about this:

In my two examples, the probability of making the right design decision was vastly improved when given the tiniest amount of empirical data.

If there’s one thing we should all take to heart, it’s that humans are strange: They rarely behave the way we expect (or want) them to. Testing often reveals issues we would never have found out by merely thinking about a design. Conversely, something that looks wrong might actually work perfectly