When I design multilingual AI assistants, I assume translation is the easy part. Intent is the hard part. A "yes" can mean "I heard you" in one culture and "I agree" in another, so I ask for confirmation in plain language. I keep a small locale playbook for tone, formality, and formats like dates and currency. Then I test with native speakers, and I read chat logs weekly. The best fixes come from real slips, like a polite reply that sounds sarcastic after translation. Messy input is the hardest part. People mix languages, abbreviate, and use slang. The upside is scale. A single agent can support multiple regions if you store language and country preferences and keep prompts local. For high-stakes steps, I add "did you mean X or Y?" before acting. This work changed naming and UX for my team. We test names for pronunciation, plus RTL and accessibility.