Building for the GCC: what multilingual really costs in AI products
Adding Arabic is not a translation task. Here is the honest breakdown of what trilingual English, Thai and Arabic support cost us in design, engineering, evaluation and content.
Every proposal we read that promises multilingual support prices it as a content line item. That estimate is wrong by roughly an order of magnitude, and the gap is almost never in translation. Having now shipped English, Thai and Arabic together in the same product, here is where the cost actually sits.
Layout is not mirroring
Right-to-left support is not a CSS flag. Logical properties get you most of the way, but the remainder is judgement: icons that indicate direction must flip while brand marks and media controls must not, numerals and Latin product names sit left-to-right inside an RTL sentence, and a progress indicator that reads naturally left to right in English needs rethinking rather than reversing. Budget design time, not just engineering time.
Typography is three separate problems
Thai has no spaces between words and taller ascender and descender stacks, so line height that looks generous in Latin looks cramped in Thai and wrong in Arabic. Arabic script needs more vertical room and a different optical weight to feel equal alongside a Latin heading. We ended up with per-script font stacks and per-script line-height and letter-spacing tokens, because one typographic scale across three scripts always makes two of them look like an afterthought.
Evaluation multiplies, it does not add
This is the cost that surprises teams. Every quality gate, safety classifier, refusal test and tone check has to run per language, because model behaviour genuinely differs across them. Our refusal rate on identical adversarial prompts varied meaningfully between English and Thai in early testing. A trilingual product does not have one evaluation suite — it has three, and they need to be maintained in parallel forever.
- Safety classifiers tuned and validated per language, not translated thresholds
- Tone and formality calibrated per culture — Thai politeness registers have no clean English equivalent
- Native-speaker review in the release process, not a QA pass at the end
- Search, sorting and text input tested per script, including keyboard and voice entry
The part that is worth it
For a company operating between the Gulf and Southeast Asia, this is not a nice-to-have. In our experience, an Arabic experience that was clearly built rather than mirrored is one of the most reliable signals of seriousness a partner in the region can look for — and it is noticed immediately. The same is true of Thai. Doing it properly is expensive once. Doing it badly is expensive every month afterwards.
Price multilingual support as three products sharing a codebase, not one product with three content files. The teams that estimate it the second way are the ones who ship two good languages and one embarrassing one.