How do you balance useful AI features with API costs?

While building an AI-powered product, I quickly realized that adding an AI feature isn't just a question of whether it works-

that's also the question of whether it makes sense to run it every time.

some AI features can genuinely improve the experience, but using AI too frequently can make the product unnecessarily expensive, especially when you're still building and don't have predictable usage.

For other makers building AI products, how have you balanced useful AI functionality with API costs?

Did you optimize prompts,reduce unnecessory calls,use different models for different tasks, or change the product design itself?

I'm curious what approaches have actually worked for you.