Chatbot development — Nepal

Chatbots and assistantsthat answer honestly

Soft Himalaya builds website assistants, WhatsApp Business bots and knowledge-based support chatbots from Kathmandu for businesses in Nepal, the UK, Australia, the USA and Canada.

  • Built around the questions your team already repeats
  • Answers grounded in content you have approved
  • A route to a person on every flow
See an example
Support team working together in an office

Support starts with people

An assistant handles agreed repeat questions and hands uncertain or sensitive cases to your team.

Support workflow photograph.
Scoped intents
We agree which questions the bot handles and which it never attempts before any flow is built.
Grounded answers
Replies are drawn from your approved content, so the bot is not inventing policy or prices.
Handover built in
Every flow has a route to a person, including outside working hours, so nobody gets stuck in a loop.
Reviewable transcripts
Conversations are logged so you can see what was asked, what was answered and what needs fixing.

What we build

Assistants for the channels you already use

Most businesses need one channel done properly rather than a bot on every surface. We scope the channel, the questions it answers and where it hands over.

Website assistant

An embedded chat widget answering questions from your own pages, documents and product information.

WhatsApp Business API

A bot on your business number using the official Meta Cloud API, within Meta's template and opt-in rules.

Messenger and Instagram

Automated replies to common direct messages, with escalation to your team for anything unusual.

Knowledge retrieval

Answers retrieved from your documents so responses cite your material rather than a model's general knowledge.

Lead capture and qualification

Structured questions that collect contact details and requirements, then pass them to your CRM or inbox.

Bookings and order status

Appointment slots, order lookups and status updates connected to the system that holds the real data.

CRM and helpdesk integration

Conversations and contacts synced to the tools your team already works in, with ticket creation where needed.

Nepali and multilingual

Handling of Devanagari, Romanized Nepali and English, tested with the phrasing your customers actually use.

Reporting and tuning

Volumes, unanswered questions and handover rates reviewed so the bot improves after launch.

What to expect

Where a chatbot helps, and where it does not

A chatbot is good at repetition and availability. It is not a replacement for judgement, and a badly scoped one costs more goodwill than it saves in staffing.

  • Repetition is the win

    Opening hours, delivery times, order status, pricing basics — the questions your team answers dozens of times a week are the ones worth automating first.

  • Grounding prevents invention

    A language model asked about your business without your content will guess. Answers are retrieved from approved material, and unanswerable questions are handed over rather than fabricated.

  • Handover is not failure

    The measure of a good bot is not how few conversations reach a human, but how few people leave without an answer.

  • It needs maintenance

    Products, prices and policies change. Without periodic review of transcripts and content, a bot slowly starts giving out-of-date answers.

How we work

From real questions to a working assistant

Timelines depend on the channel, how ready your content is and which systems the bot needs to read from. We plan the schedule once the intents are agreed.

  1. 01

    Intent discovery

    We review your real enquiries — inbox, chat logs, phone notes — and list the questions the bot should handle, and the ones it should not.

  2. 02

    Content and data sources

    The material each answer will be drawn from is identified: pages, documents, product data or a system the bot needs to query.

  3. 03

    Flow and tone design

    Conversation paths, fallback wording, handover rules and response tone are drafted and approved before implementation.

  4. 04

    Build and integration

    The assistant is built on the chosen channel and connected to the CRM, helpdesk or business systems in scope.

  5. 05

    Testing with real phrasing

    Flows are tested with the way customers actually write, including mixed languages, typos and questions outside the scope.

  6. 06

    Launch and review

    The bot goes live with logging enabled, and transcripts are reviewed on an agreed cycle to correct gaps and wrong answers.

Example

How a well-scoped bot behaves

This is a scripted walkthrough rather than a live assistant. The interesting cases are the last three: what a bot does when it should not answer matters more than what it does when it can.

Try a question

Support assistantScripted demonstration
  • Hello. This is a scripted demonstration — pick a question below to see how the assistant is designed to respond.

Your assistant would answer from your own content and connect to your own systems. This demo does not send anything anywhere.

Tools

What we build with

Model and platform choices are made per project. Where a paid API is involved, the account is set up in your name so usage and cost stay visible.

  • Claude and OpenAI APIs

    Language models for open-ended conversation

  • Meta WhatsApp Cloud API

    Official WhatsApp Business messaging

  • Retrieval over your content

    Answers grounded in your documents and pages

  • Node.js and Next.js

    Webhooks, backends and embedded widgets

  • PostgreSQL and vector search

    Conversation storage and content lookup

  • CRM and helpdesk APIs

    Lead sync and ticket creation

Where it fits

Common situations we are asked about

These are the patterns businesses usually bring us. Each still starts with the same question: which enquiries repeat often enough to be worth automating?

Retail and e-commerce

Stock and delivery questions, order status lookups and product guidance during browsing hours and after them.

Education

Admission criteria, fee structures, deadlines and document requirements answered from published information.

Clinics and services

Appointment availability and preparation instructions, with anything clinical routed to qualified staff rather than answered.

Travel and hospitality

Package details, availability enquiries and pre-booking questions, with the booking itself handled by your system.

Property and services

Listing details, viewing requests and qualification questions passed to the agent handling that enquiry.

Logistics and support

Consignment tracking, common troubleshooting steps and ticket creation when the issue needs a person.

Project estimates

Scope first, then price

Cost depends on the channel, how many intents are in scope, whether the bot reads from live systems and how much content preparation is needed. Model and messaging usage is billed by the platform and is separate from our fee.

Scripted assistant

A defined set of questions and flows — hours, services, directions, simple lead capture.

  • Agreed intent list
  • Fixed conversation flows
  • Lead capture to email or CRM
  • One channel

Knowledge-based assistant

Answers retrieved from your own content, with handover rules and transcript review.

  • Content preparation
  • Retrieval-grounded answers
  • Handover and escalation
  • Post-launch tuning

Connected assistant

A bot that reads from or writes to your live systems — orders, bookings, accounts or tickets.

  • System integration design
  • Authentication and data rules
  • Failure and fallback handling
  • Monitoring and review cycle

Questions

Frequently asked questions

Common questions about channels, AI versus scripted flows, data handling and how chatbot work is quoted.

It is the work of deciding which questions an automated assistant should handle, where its answers come from, and what happens when it cannot help. That covers scripted flows, language-model assistants grounded in your own content, the channel it runs on, and the connections to your CRM or support tools.

Working together

A clear chatbot relationship

We do not publish anonymous client quotes, satisfaction scores or deflection percentages that cannot be checked. These are the standards we hold ourselves to instead.

Scope agreed in writing

The intents, channels, integrations, handover rules and review cycle are documented before development starts.

You own the accounts

Model API keys, WhatsApp Business accounts and any platform subscriptions are created in your name.

Answers you approved

The bot responds from content you have signed off, and unanswerable questions go to a person rather than a guess.

Honest measurement

We report conversation volumes, handover rates and the questions the bot could not answer, not just the flattering totals.

Start a project

Tell us which questionsyou keep answering

Send the enquiries your team repeats most, the channel you want the bot on and any system it would need to read from. We will reply with questions, a suggested scope and an estimate.

This opens your email client with the details filled in. Please do not include sensitive information.