Why Your Chatbot Needs to Understand Romanized Nepali

Open the inbox of any Nepali business page and read the messages. You'll find very little textbook English and not much Devanagari. What you'll mostly see looks like this:

• "price kati ho?"
• "M size cha?"
• "delivery kahile samma aaucha Biratnagar ma"
• "esewa bata pay garna milcha?"

This is Romanized Nepali: Nepali written with English letters, often mixed with English words. It's how most people in Nepal type, because it's fast and works on any keyboard.

It's also where most chatbots fall over.

Why it breaks most chatbots

Keyword-based bots look for exact words. "Price" might work, but "kati ho", "rate", "kati parcha" and "kati" all mean the same thing, and the bot has to know every one. Add spelling variations (cha, xa, chha) and the list never ends.

Many AI chatbots built for international markets struggle too. They're tuned for English, so a Romanized Nepali message either gets misunderstood or gets an awkward, overly formal English reply that feels foreign.

The customer doesn't care why. They just see a page that didn't understand them, and they message the competitor instead.

What good looks like

A chatbot that works for Nepali customers should:

1. Understand all three styles: English, Devanagari Nepali and Romanized Nepali, including mixed messages.
2. Handle spelling variations without you listing them: "cha", "xa" and "chha" are the same word.
3. Reply in the customer's style. If they write in Romanized Nepali, a short, natural Romanized reply usually feels better than formal English.
4. Know local context: places like Lalitpur or Dharan, payment methods like eSewa, Khalti and Fonepay, and cash on delivery.

How to test a chatbot before you buy

Don't trust a demo script. Send it the messages your customers actually send. Try a few of these:

• "yo kurta ko price kati ho? L size cha?"
• "Pokhara ma delivery charge kati lagcha?"
• "COD milcha ki advance tirna parcha?"
• "kaile samma aaucha? urgent chaiyeko"
• A message with a spelling mistake, or half Nepali and half English

Then check two things: did it understand the question correctly, and does the reply sound like something your own staff would say?

Why romanized text is hard for software

Romanized Nepali has no standard spelling. The same word can be written three or four ways, and people mix in English freely, sometimes in the middle of a sentence. A single message might contain Nepali grammar, English nouns and a typo or two:

• "sir yo phone ko price kati ho, emi ma milcha?"
• "kal sma deliver hunxa? urgent xa"

For a keyword bot, that's chaos. For a modern AI model it's manageable, but only if the system is set up to expect it and tested on it, rather than assuming customers write tidy English.

The same is true far beyond Nepal. Hindi and Urdu speakers often type in Roman script, and so do many Arabic speakers. Any business whose customers write the way they speak runs into the same problem.

Setting up your bot for mixed-language customers

1. Collect real examples. Copy 30 to 50 actual customer messages from your inbox, typos and all.
2. Test with them before launch. Every one should be understood correctly.
3. Decide your reply style. Match the customer's language, or reply in English unless they write in Devanagari. Either works; being consistent matters.
4. Keep product names consistent. If customers call it "jhola" and your sheet says "tote bag", add both.
5. Review weekly. New slang and spellings appear all the time, especially around festivals and trends.

It isn't only about Nepali

None of this is specific to one language. Orbit replies in whatever language a customer writes, whether that's English, Hindi, Arabic, Spanish or Nepali, and handles romanized and mixed text the same way. Romanized Nepali is simply the clearest example of why "supports English" isn't good enough for a business that sells to real people.

Why this matters for your sales

Language isn't a small detail here. When customers feel understood, they keep talking, and the longer the conversation goes the more likely it ends in an order. When they get a confused reply, the conversation usually ends right there.

Orbit was built in Kathmandu with exactly this in mind. It replies in whatever language a customer writes, Nepali and Romanized Nepali included, and matches their style, on Messenger, Instagram, WhatsApp, TikTok and your website. Book a demo and test it with your own customers' messages.