what's your use case? Because LLMs can count, they're just not as consistent at it as straight code it. For best results, I'd recommend doing your precentages in code and then pass the value back to the AI