When should we actually consider dynamic scaling beyond basic auto-scaling groups?

Hey everyone, I've been thinking about how we scale things, especially for systems where traffic is all over the place or resources are super specific. Usually, we just use basic auto-scaling groups, you know, with simple CPU or memory triggers, and that's totally fine for a lot of stuff. But when do we really need to look at smarter ways to scale? I mean like predictive scaling, or when we scale based on how long a message queue is, or even using custom info from our monitoring tools. Is it just about saving money, or making sure performance is good when things get busy, or what? I'm curious to hear what you all have done and when you decide to go past the usual setup.