Most chatbot projects in Nepal fail for a simple reason. The bot is built for clean English, while customers write the way they actually speak. One person types “मूल्य कति हो?”, another types “mulya kati ho?”, and a third types “price kati ho?” All three mean the same thing, but a weak system treats them as separate problems.
If your team is planning a support bot, a lead capture bot, or a FAQ bot, language support is not a side feature. It is the job.
Start with the job before the model
A support bot, a lead bot, and a FAQ bot are three different systems. They need different conversation flows and different checks.
- A support bot answers repeated questions and moves hard cases to an agent.
- A lead bot collects contact details, asks short follow-up questions, and passes clean leads to sales.
- A FAQ bot stays inside approved content and should not guess outside it.
If one bot tries to do all three jobs on day one, the flow usually gets muddy fast.
Match the way users type
Customers in Nepal do not stay inside one script. They move across formats depending on keyboard, mood, and speed. A working bot should accept at least these forms.
- Devanagari, such as “डेलिभरी कति बजे हुन्छ?”
- Romanized Nepali, such as “delivery kati baje huncha?”
- Mixed chat, such as “refund pauncha if size milena?”
- Short mobile messages with spelling drift.
The reply should usually follow the user's style. If the customer starts in Devanagari, answer in Devanagari. If the customer writes short Romanized Nepali, answer in a style that feels natural in that same thread.
Do not force the user to match the bot. Train the bot to match the user.
Code-switching is normal
People mix English into Nepali all the time because product names, payment terms, delivery steps, and support actions often show up that way in daily chat. The model should not treat that as broken language.
A delivery intent might show up as:
- “Delivery kati baje samma huncha?”
- “delivery time k ho?”
- “aaja delivery possible?”
- “आज डेलिभरी हुन्छ?”
Those variations should point to the same answer path. This is where real data beats generic examples every time.
Devanagari support is more than text rendering
Some teams think Nepali support is done once the font looks right on screen. That is only the first check. A good Devanagari experience also needs stable input handling, search that survives spelling variation, and replies that sound like business chat, not machine-translated text.
Input handling
Users should be able to paste or type Nepali without characters breaking or changing after send.
Search and matching
The bot should still find the right answer when punctuation is missing, a place name is written a little differently, or the user shortens the phrase.
Output tone
Replies in Nepali should stay short and direct. Long, stiff answers are slow to read on mobile and easy to skip.
Romanized Nepali needs its own treatment
Romanized Nepali is messy by nature. The same word can appear as “cha”, “chha”, or “xa”. The same phrase can show up in three or four shapes. Customers understand that. Your bot should too.
You do not need one national spelling standard inside the product. You do need a normalization plan. Build a list of common variants, slang, and abbreviations from your own chats. A clinic, a school, and an ecommerce store will all see different patterns.
Generic Nepali examples are a weak stand-in for the language your own customers already use.
Build the dataset before you write prompts
Most of the work sits in data prep. Start with material that already exists inside the company.
- Website FAQs
- Instagram and Facebook inbox chats
- WhatsApp support messages
- Email support threads
- Call-center notes
- Store policies, delivery rules, refund rules, and branch details
Then clean it. Remove old policies. Remove duplicate answers that say different things. Fix any answer that depends on staff memory instead of a written rule.
- Name the intent.
- Collect sample user questions in different scripts.
- Write one approved answer.
- List the follow-up questions the bot can ask.
- Mark the cases that must go to a human.
Keep answers tight
Mobile chat rewards short answers. A good support reply answers the question first, then asks for one missing detail if needed. A good lead bot asks one thing at a time. If the user wants solar installation, do not ask for name, phone, address, budget, and roof size in one block. Start with the phone number, then the location, then the next qualifying detail.
Test with real Nepali chat, not only clean test prompts
A bot can look good in an internal demo and still fail live. Your test set should include angry messages, short messages, spelling mistakes, follow-up questions with no context, mixed Nepali and English, and messages that clearly ask for an agent.
- If a user writes “Kathmandu bahira delivery cha?”, the bot should not answer with valley rates.
- If a user writes “Yo stock ma cha?”, the bot should ask which item unless the product context is already known.
- If a user says “agent sanga kura garna paryo”, the bot should hand off.
- If a user says “mero payment katyo tara order dekhaena”, that is a support case, not a FAQ answer.
Set handoff rules before launch
A good bot is not one that answers everything. A good bot knows when to stop. Put the handoff rules in writing before the chatbot goes live.
- Payment problems and refund disputes.
- Order changes after confirmation.
- Medical, legal, or financial advice.
- Repeat failure to understand the user.
- Any direct request for a human.
The handoff should carry context. The agent should receive the chat transcript, detected intent, collected customer details, and the reason the bot stepped aside.
Roll out in stages
Do not launch on every channel at once. Start with one use case and one channel. For many teams, the safest first launch is a website FAQ bot with approved answers only. After that, add a support flow with human handoff. Then add lead capture. Then expand to WhatsApp, Messenger, or Instagram.
Read failed chats every week. Add missing phrasing. Remove weak answers. Tighten handoff rules. If the same question breaks three times in one week, fix it that week.
Measure the right things
- Correct answer rate.
- Handoff rate.
- Repeat contact rate after the bot reply.
- Lead completion rate for lead bots.
- Questions with no useful match.
Numbers matter, but chat logs matter too. The logs show tone problems, policy confusion, and language mismatches that a dashboard can miss.
The standard to aim for
A good Nepali chatbot sounds like your team on a good day. It is clear, polite, short, and accurate. It keeps up when the customer changes script halfway through the message. It understands enough Romanized Nepali to stay useful. It stops when a human should take over.
Nepal AI Hub builds chatbot systems around that standard, with support for Nepali language input, clean escalation, and workflows that fit the way local teams already handle customer questions. If you want to scope a chatbot for your business, start with the live conversations you already have. They will tell you what the bot really needs to learn.