Disclosure: I'm the founder of GulfBot. This post is about product and UX decisions, not a stack tour.

GulfBot is an AI assistant plus a service marketplace for GCC business tasks: company setup, Golden Visa, PRO and VAT, across the UAE, Saudi Arabia, Qatar, Kuwait, Bahrain and Oman. Our users are an unusually mixed crowd: Emirati and Saudi business owners, Indian and Keralite founders, Latin American investors, and expats who've lived in Dubai for twenty years. So we ship in five languages: English, Arabic, Hindi, Spanish and Malayalam.

Here are the decisions that mattered most, and what I'd tell anyone building something similar.

1. Treat RTL as a layout mode, not a translation

Arabic isn't "English text, but Arabic". When the language flips, the whole interface mirrors: navigation order, back arrows, progress steps, chat bubble alignment, icons that imply direction. The mistakes we made early were small and everywhere: a "next" chevron pointing the wrong way, numbers and English brand names inside Arabic sentences rendering in a confusing order, padding that only looked right left-to-right.

Lesson: design every screen once in an RTL mock before you call it done. Test mixed-direction strings ("GulfBot", amounts in AED, document numbers inside Arabic text) on purpose.

2. Let people switch language anywhere, and remember it

A lot of our users think in one language and do paperwork in another. Someone might read service details in Malayalam and then want the official term in English to compare with an agent's quote. So switching is always one tap away, it doesn't reset the conversation or the request they're working on, and the app remembers their choice.

Lesson: language is a user preference, not an onboarding step. Don't bury it in settings and don't punish switching.

3. Plain-language answers beat "complete" answers

Business setup is full of jargon: MoA, establishment card, qualifying income, PRO. The natural instinct with an AI assistant is to be thorough. For our users, thorough meant unreadable. We push for short answers that say what it means for you, then the next step, and only then the detail.

This matters even more across languages. A literal translation of a dense legal sentence into Hindi or Malayalam is technically correct and practically useless.

Lesson: write a style rule for the assistant ("explain like you're talking to a smart friend who has never set up a company") and check it in every language, not just English.

4. Keep official terms recognizable

The flip side of plain language: some terms must stay recognizable, because users will see them on government portals and invoices. "Golden Visa", "free zone", "trade licence", "VAT". We explain them in the user's language but keep the official term visible alongside it.

Lesson: maintain a small glossary of terms that should never be fully localized away.

5. Be honest about what the AI is

Our assistant is informational. Approvals come from authorities, and rules differ by emirate and free zone. So the product says so, in every language, and every answer leads to an actual service you can request and track (status, documents, next steps) rather than leaving the user with a chat transcript.

Lesson: in regulated domains, the AI's job is to get people to the right next action with the right questions, not to pretend to be the final word. A disclaimer that only exists in English isn't a disclaimer.

What I'd do differently

  • Start with RTL and one non-Latin script from day one. Retrofitting is painful.
  • Get native speakers to review real conversations, not just UI strings.
  • Track which language users switch to mid-task. It tells you where your content is weakest.

If you're building multilingual products, especially for the Gulf or South Asia, I'd love to hear what tripped you up. Drop it in the comments.

You can see how it turned out at https://gulfbot.app/?utm_source=devto&utm_medium=social&utm_campaign=oct2026