Getting good AI results in Bahasa Indonesia

Most large language models are trained mostly on English. They speak Bahasa Indonesia well enough to demo — and just badly enough to embarrass you in production if you’re not careful. Here’s what actually moves quality.

Test in the language, not in English

The most common mistake: teams evaluate a model in English, ship it, and discover the Indonesian output is stiff, over-formal, or subtly wrong. Always test on real Indonesian inputs — including the messy, mixed-language, slang-filled way customers actually write on WhatsApp.

Watch the register

Indonesian carries formality that English doesn’t. A model defaulting to textbook Bahasa baku can sound cold to a customer expecting warmth, or too casual for a formal notice. Pin the tone deliberately in the system prompt, with examples.

Ground it in your own words

Retrieval (RAG) over your Indonesian documents beats relying on the model’s general knowledge — it keeps terminology, product names, and phrasing consistent with how you speak, not how the internet does.

Consider the mix

Real Indonesian business writing code-switches — English tech terms inside Indonesian sentences. A good setup handles that gracefully instead of “correcting” it into something no one would say.

When bigger isn’t better

Sometimes a model with stronger multilingual training beats a bigger, English-centric one on your task. Test both. Vendor-neutral means we pick the one that scores best on your Indonesian, not the one with the loudest marketing.

Language is where a lot of AI projects quietly lose trust. Get it right and everything downstream feels more reliable.