AI Chatbot & Agent Development in Nepal
Your WhatsApp gets the same thirty questions every day: the price, the opening hours, do you deliver to Lalitpur, is the doctor in on Saturday. This is the work that builds a chatbot or an agent that answers those from your own documents, and hands the rest to a person.
Animated chat conversation: a visitor's message goes to an AI agent core, the agent calls an external tool such as a booking calendar, receives the result and replies in the chat, showing AI chatbots and agents as assistants that can act as well as answer.
I'm Shashikant Shah, a digital marketing expert in Kathmandu and a software engineer. Chatbot and agent development is software work: a bot on your website or WhatsApp that answers from your own documents and policies, and, where it's worth it, an agent that can take an action like booking a slot or creating a ticket. It's for businesses in Nepal fielding the same questions all day. It sits next to my n8n automation work and under the AI automation expert page.
Short answer: AI chatbot and agent development in Nepal means a bot that retrieves answers from your own price list, FAQs and policies instead of guessing, works on your website and WhatsApp, hands over to a person the moment it's unsure, and is watched after launch so wrong answers get fixed. Build takes four to eight weeks, and you keep the code and the accounts.
When you need this
A trekking company in Thamel answers "how much is the Everest Base Camp trek" forty times a day on WhatsApp, at midnight for customers in Australia and the UK. A hospital reception in Kathmandu picks up the same call about OPD hours. A shop selling on Daraz and Instagram gets "is this available in size M" and answers three hours later, after the sale has gone elsewhere.
The generic widget doesn't fix this. A model answering from its general training will confidently invent your refund policy or your consultation fee. The bots I see installed in Nepal are either that, or a menu of buttons that sends everyone to "talk to an agent" anyway.
And some things shouldn't be automated at all. A complaint, a medical question, a price negotiation with a dealer: the bot's job there is to recognise it fast and hand it to a person with the context attached.
The problems I see most often:
- A chat widget that makes up answers because it was never connected to your real documents.
- WhatsApp replies that go out hours later, when the customer messaged three businesses at once.
- A bot that can't say "I don't know", so it guesses at prices and opening hours.
- No handover, so the customer repeats everything to a person after fighting the bot.
- Nobody reading the conversations after launch, so the same wrong answer goes out for months.
What I actually do
- Scope: which questions the bot answers, which it hands over and which it never touches, drawn from your real WhatsApp and call history, so it's built for the thirty questions you actually get. You get a bot with a job description.
- Retrieval over your own documents: price lists, FAQs, policies, product sheets and pages from your site, kept in a knowledge base you can update yourself. The bot answers from those and says so. You get answers that are true about your business.
- Website and WhatsApp channels: a chat on your site, and the same bot on WhatsApp Business where that's how your customers talk, with Viber and Messenger where they matter. You get one bot, not three.
- Agents with tools, where it's worth it: booking a slot, creating a support ticket, updating a lead in your CRM or checking stock, through the same n8n workflows I build. Anything with consequences waits for a person. You get actions, with a checkpoint.
- Human handover: the bot recognises a complaint, a medical or legal question, or a customer going in circles, and passes it to a person with the conversation attached. You get fewer angry customers, not more.
- Monitoring after launch: I read the conversations, log the wrong or missing answers and fix the knowledge base, weekly at first. Hospitals and clinics can see how it fits on the healthcare page. You get a bot that improves instead of drifting.
- Tracking: enquiries, handovers and bookings counted in your own Analytics through the tracking work, so you can see what the bot handled and what it cost.
How it works
First call and scoping. The first thirty minutes are free. You show me the questions you get most, on WhatsApp, the phone and your inbox, and I tell you what a bot can handle, what it shouldn't, and whether it's worth building at all. If you go ahead, you get a written scope: the question list, the documents it needs, the handover rules, the channels and what's not included.
The work. Four to eight weeks, on your accounts: your OpenAI or Claude API key, your WhatsApp Business number, your hosting. I build the knowledge base first and test the bot against the real questions with you before it meets a customer. Then it goes live on one channel, with handover on from day one, and a second channel once the answers hold.
Reporting, and what you keep. After launch you get a monthly note on conversations, handovers, the questions it couldn't answer and the API cost. You keep the code, the repository, the knowledge base, the API keys and the hosting, all in your name. If we stop, the bot keeps running without me.
Proof from AI chatbots and agents work
I haven't published a chatbot or agent result with numbers yet, and I won't invent one. No named client on my case studies page has a chatbot in the scope I delivered, and I'd rather say that than dress up a demo as a result.
The nearest work by mechanism is the automated publishing I delivered for Goenka Gold, Red Mud Coffee, Shree Ridhi Sidhi Jewellers and FoodMario, which is the same idea: a repeated job handled by a workflow, with a person checking what matters. This site is the other evidence of how I build: it passes 15 automated checks on every deploy, and I wrote it myself. The difference between an agent and a plain automation is in my article AI agents vs traditional automation.
What AI chatbots and agents costs
A chatbot or agent is quoted per project after the first call, because the scope decides the price. Everything I build is created in your name, on accounts and hosting you control, so ending the engagement costs you nothing but my access. My published retainer for ongoing work is USD 1,800 a month, about NPR 245,000, and the audit and sprint prices are on the pricing page. The first thirty minute call is free, and if a WhatsApp Business auto reply and a good FAQ page would do the same job for you, I'll say so and you won't need me.
Questions about AI chatbots and agents
Is this different from a chatbot widget I could install myself?
Yes. A widget answers from the model's general training, which means it guesses at your prices and policies. What I build retrieves from your own documents, hands over to a person when it's unsure, and gets read and corrected after launch. If a simple button menu would do the job, I'll tell you to install that instead.
Can the agent take real actions, not only answer questions?
Yes, within an agreed list: booking a slot, creating a ticket, updating a lead in your CRM, checking stock. Anything with consequences, like a payment or a cancellation, waits for a person to approve. The list is in the written scope before I build anything.
Does it work on WhatsApp, and in Nepali?
Yes to WhatsApp, through the WhatsApp Business platform on your own number. Nepali, English and the mix people actually type can all be handled, but only if your documents and the test questions include them, so I test with your real messages before launch. That's where most bots in Nepal fall over.
How long does it take?
Four to eight weeks for a website or WhatsApp bot with retrieval and handover. The knowledge base is most of the work: if your prices and policies live only in people's heads, writing them down comes first and adds time. An agent that takes actions adds two to four weeks for the integrations and the approval steps.
What does it cost?
Per project, after the first call, because the channels, the documents and the actions decide the price. You pay the API usage and the WhatsApp fees directly, so there's no markup. My published retainer for monitoring and improvement is USD 1,800 a month (about NPR 245,000 a month), and the pricing page has the rest.
When are you not the right person?
If you want a bot to handle complaints, medical advice or anything where a wrong answer hurts someone, no. Those go to a person, and I build the handover, not the replacement. If you get five questions a day, a good FAQ page and WhatsApp quick replies are cheaper, and I'll say so.
Pages related to AI chatbots and agents
All services · AI automation with n8n · CRM development · AI automation expert in Nepal · AI automation for business in Nepal · AI agents vs traditional automation · Pricing.
Tell me about the questions you get
Send me the ten questions your WhatsApp gets most. I'll tell you which a bot should answer, which need a person, and whether it's worth building at all, whether or not you hire me. The first thirty minutes are free.
