AI chatbot development is the process of building a conversational assistant, usually powered by a large language model, that answers customer or staff questions on your website, WhatsApp or internal tools, grounded in your own information and connected to your systems. In 2026 a well-built business chatbot typically costs about $5,000 to $15,000 for a document-based FAQ bot, $15,000 to $40,000 for one connected to your systems, and more for multi-channel, transactional bots, plus monthly running costs from a few dollars to a few hundred.
The technology is no longer the hard part. Any developer can connect a model to a chat window in a day. What separates a useful chatbot from an embarrassing one is grounding (answering only from approved information), knowing when to hand over to a person, and measuring results honestly.
This guide covers the full picture for a general business: when a chatbot is the right tool, the development process step by step, the tech stack, website vs WhatsApp, cost to build and run in USD, INR and AED, and the mistakes to avoid. Hospitals and clinics should also read our healthcare-specific guide, AI chatbot for hospitals.
Chatbot, AI assistant or AI agent: which do you need?
| Type | What it does | Example | Build effort |
|---|---|---|---|
| Rule-based chatbot | Follows fixed menus and buttons | "Press 1 for store hours" | Low |
| AI FAQ chatbot | Answers free-text questions from your documents | Policy, product and pricing questions | Low to medium |
| Connected AI chatbot | Answers using live data from your systems | Order status, appointment availability | Medium |
| Transactional chatbot | Completes simple tasks | Booking, rescheduling, lead capture | Medium to high |
| AI agent | Decides steps and takes actions across systems | Resolving a delivery issue end to end | High |
Most businesses should start in rows two or three. If you need the last row, read about AI agent development before you scope it.
The AI chatbot development process, step by step
- Pick the questions that matter. Export the last three to six months of enquiries from email, chat, WhatsApp and phone notes. Group them. Usually 20 to 40 question types cover most of the volume.
- Decide scope and hand-off rules. List which questions the bot answers, which it hands to a person, and how quickly a person responds in office hours and outside them.
- Collect and clean the knowledge. Policies, product information, price lists, FAQs. Remove outdated versions; this step alone often fixes half the wrong answers.
- Design the conversation. Greeting, tone, how the bot asks for missing details, what it says when it does not know, and how it confirms a booking or lead.
- Build retrieval and integrations. Index the knowledge so the bot can search it, and connect to systems such as your CRM, booking calendar or order database through APIs.
- Write the instructions and guardrails. Answer only from retrieved information, cite sources where useful, never invent prices or policies, and hand over when unsure.
- Test with real questions. Build a test set from the enquiries you exported, including off-topic, rude and tricky questions, and score the answers.
- Soft launch. Put the bot on one channel or one page first, with staff reviewing conversations daily.
- Measure and improve. Track resolution, hand-off and satisfaction, and fix the top failure each week.
Grounding: why your chatbot needs RAG
A raw language model answers from its general training, which means it will confidently describe a return policy you do not have. Retrieval-augmented generation (RAG) fixes this: for each question, the system searches your approved documents, passes the relevant passages to the model and instructs it to answer only from them. Our explainer on what RAG is covers how it works, and RAG vs fine-tuning explains why RAG, not fine-tuning, is the right default for business chatbots.
For questions that depend on live data, such as "is my order shipped?", the bot calls your systems through APIs instead of searching documents.
The AI chatbot tech stack in 2026
| Layer | Common choices | Notes |
|---|---|---|
| Language model | OpenAI GPT models, Anthropic Claude, Google Gemini, open-weight models such as Llama or Qwen | Test two or three on your questions; smaller models often suffice |
| Retrieval | Vector search in PostgreSQL (pgvector), managed vector databases, or hybrid keyword plus vector search | Hybrid search handles product codes and names better |
| Backend | Node.js or Python service handling conversations, tools and logging | Keep API keys and business rules here, not in the browser |
| Channels | Website widget, WhatsApp Business Platform, Instagram, Slack or Teams for internal bots | Each channel has its own formatting and rules |
| Automation | n8n, Make or custom code for follow-ups and CRM updates | See workflow automation |
| Monitoring | Conversation logs, evaluation runs, cost tracking | Non-negotiable for customer-facing bots |
Website chatbot vs WhatsApp chatbot
In India and the UAE, WhatsApp is often where customers already are; in the US and UK, website chat and email still dominate many sectors. Many businesses need both, built on one backend.
| Website chatbot | WhatsApp chatbot | |
|---|---|---|
| Where customers are | Visitors already on your site | Customers on their phones |
| Setup | Embed a widget | WhatsApp Business Platform via Meta or a provider, verified business, approved templates |
| Messaging costs | None beyond hosting | Meta charges per template message by category and country |
| Proactive messages | Only while visitor is on site | Allowed with approved templates and opt-in |
| Rich features | Full control of the interface | Buttons, lists, WhatsApp Flows, payments in some markets |
| Best for | Pre-sales questions, support, lead capture | Bookings, reminders, order updates, repeat customers |
Meta moved the WhatsApp Business Platform to per-message pricing on 1 July 2025. Replies to customer-initiated messages within the customer service window are free, utility templates sent inside that window are free, and marketing, authentication and out-of-window utility templates are charged per message at rates that vary by country (Meta's WhatsApp pricing documentation).
How much does AI chatbot development cost?
Indicative build costs, converted at roughly ₹95 and AED 3.67 to the dollar:
| Chatbot type | USD | INR | AED | Typical timeline |
|---|---|---|---|---|
| AI FAQ bot on your documents, one channel | $5k-15k | ₹4.75-14 lakh | AED 18k-55k | 3-6 weeks |
| Connected chatbot (CRM, bookings or orders), one or two channels | $15k-40k | ₹14-38 lakh | AED 55k-147k | 6-10 weeks |
| Transactional, multi-channel chatbot with hand-off inbox | $30k-80k | ₹28.5-76 lakh | AED 110k-294k | 2-4 months |
| Agent-level assistant with several systems and approvals | $60k+ | ₹57 lakh+ | AED 220k+ | 3-6 months |
India-based teams typically sit at the lower end of each range for the same scope. One Indian vendor's published estimate puts a 400-hour chatbot build at about $14,000 with an India team against about $46,000 with a US team, and quotes generative AI chatbots at $40,000 to $100,000 and more at the high end (SolGuruz, vendor-reported).
What moves the price:
- Number of integrations and how good their APIs are.
- Number of channels and languages (Arabic, Hindi, Malayalam and other regional languages need separate testing).
- Volume and quality of knowledge to clean and index.
- Whether the bot only answers or also takes actions such as bookings.
- Compliance needs, such as HIPAA for US health data, GDPR in the UK and EU, India's DPDP Act and Rules 2025, and UAE data protection law.
How much does it cost to run an AI chatbot each month?
Model usage for a chatbot handling 10,000 conversations a month, at about four turns each, with roughly 3,000 input tokens (instructions, retrieved passages, history) and 250 output tokens per turn, using Anthropic's published API prices:
| Model | Cost per turn | Per conversation | Per month (10,000 conversations) | INR | AED |
|---|---|---|---|---|---|
| Claude Sonnet 5.5 ($2 / $10 per million tokens) | $0.0085 | $0.034 | $340 | about ₹32,300 | about AED 1,250 |
| Claude Haiku 5.5 ($0.10 / $0.50 per million tokens) | $0.000425 | $0.0017 | $17 | about ₹1,600 | about AED 62 |
Prices from Anthropic's pricing page, November 2026. Prompt caching of the stable instructions lowers the mid-tier figure further. OpenAI and Google price comparably, and the same arithmetic applies.
Add to that:
- Hosting, database and vector search: often $20 to $300 a month at this scale.
- WhatsApp template fees: depend on how many proactive or out-of-window messages you send and to which countries.
- Monitoring and improvement time: a few hours a week from someone who owns the bot.
- Maintenance: updates when policies, products or models change.
Hand-off to humans: design it first, not last
The fastest way to lose customer trust is a bot that traps people. Good hand-off design:
- A visible "talk to a person" option at any point.
- Automatic hand-off on frustration, repeated failure, complaints, payments disputes or sensitive topics.
- The full conversation and a short summary passed to the agent, so the customer never repeats themselves.
- Honest expectations out of hours ("A person will reply by 10 am tomorrow").
- Tracking of hand-off reasons, which become your improvement backlog.
Our guide to AI customer support automation metrics covers how to measure deflection and quality without fooling yourself.
Common chatbot mistakes to avoid
- Launching on every channel at once.
- Letting the bot answer from general knowledge instead of your documents.
- No owner, so nobody reads conversations after week two.
- Measuring "conversations handled" instead of questions actually resolved.
- Hiding the human option to inflate deflection numbers.
- Ignoring prompt injection and abuse; a public bot will be tested by strangers.
Choosing an AI chatbot development company
Look for a partner who starts from your enquiry data rather than a template, assigns a named owner to each step, shows you the bot answering your real questions within the first couple of weeks, builds on your existing CRM and booking systems, and reuses proven components so you are not paying for everything from scratch. Time-zone overlap matters for launch weeks.
How Algon approaches this: Algon is a Kochi-based team building chatbots and assistants for clients in India, the UAE and the USA. We begin with your past enquiries, ground answers in your approved information with sources shown, design hand-off to a named person from day one, and run the bot alongside staff before widening its scope. See our LLM development services and RAG development pages for how we build them.
Conclusion: a chatbot is a knowledge and process project
AI chatbot development in 2026 is affordable and fast, but quality depends on clean knowledge, sensible scope, honest hand-off and an owner who keeps improving it. Start with the questions that fill your inbox, ground every answer, launch on one channel, and expand once the numbers show it is working.
Frequently asked questions
How much does it cost to build a chatbot?
Indicatively, an AI FAQ chatbot on your documents costs about $5,000 to $15,000, a chatbot connected to your CRM or booking system about $15,000 to $40,000, and a multi-channel transactional chatbot $30,000 to $80,000. India-based teams usually sit at the lower end.
How much does it cost to run an AI chatbot each month?
For 10,000 conversations a month, model usage can range from under $20 on a small model to around $340 on a mid-tier model at current Anthropic list prices. Add hosting, any WhatsApp template fees, and a few hours a week of review time.
How much does an AI chatbot cost in India?
With an India-based team, a document-based AI chatbot typically costs around ₹4.75 lakh to ₹14 lakh, and a connected chatbot ₹14 lakh to ₹38 lakh, depending on integrations, channels and languages. Monthly running costs can start from a few thousand rupees.
How do I build a custom chatbot for my business?
Start from your real enquiries, decide what the bot answers and what goes to a person, clean your documents, add retrieval so answers come only from approved information, connect the systems it needs, test on real questions and launch on one channel first.
Should my chatbot be on my website, WhatsApp or both?
Go where your customers already message you. In India and the UAE that is often WhatsApp; in the US and UK website chat is common. Many businesses run both on one backend so answers and hand-offs stay consistent.
Will an AI chatbot give wrong answers?
It can if it answers from general knowledge or outdated documents. Grounding answers in approved, current information with retrieval, instructing it to hand over when unsure, and reviewing conversations regularly keeps wrong answers rare.
