Ask three agencies what an AI chatbot costs and you will get three wildly different numbers — anywhere from ₹40,000 to ₹25 lakh. That is not because someone is lying; it is because "AI chatbot" covers everything from a scripted FAQ widget to a retrieval-augmented assistant wired into your live database. The word "chatbot" hides a 60x price range. This guide breaks down exactly what drives the price in 2026, gives you realistic India-market ranges, walks through a worked budget, and shows you where the money quietly leaks after launch.
First, what kind of chatbot are you actually buying?
Most projects fall into one of three buckets. The tier you need is decided by one question: does the bot need to understand language and use your data, or just follow a script? Pricing below reflects typical Indian agency and senior-freelance rates in 2026.
| Tier | What it is | Typical build cost (INR) | Best for |
|---|---|---|---|
| Rule-based / flow bot | Scripted decision trees, button flows, no real language understanding | ₹40,000 – ₹1,50,000 | Lead capture, FAQ deflection, appointment booking |
| AI / RAG chatbot | LLM-powered, answers from your own docs and data via retrieval, remembers context | ₹1,50,000 – ₹6,00,000 | Customer support, internal knowledge, product Q&A |
| Custom / agentic assistant | Multi-step tool use, takes actions in your systems, fine-tuning, guardrails | ₹6,00,000 – ₹25,00,000+ | Complex workflows, regulated domains, scale |
The jump from tier one to tier two is the expensive one, and it surprises people. A flow bot is essentially UI work — you are paying for screens and logic. A RAG chatbot is a small data-and-retrieval system that has to be built and evaluated, because the moment a bot generates answers instead of reading them off a script, "is it correct?" becomes a real, ongoing engineering question.
What actually drives the number
Two bots with the same tier can differ 3x in price. Here is where the hours really go:
- Integration surface. A bot that answers from a PDF is cheap. A bot that reads your live orders, checks stock, cancels a booking, or writes to your CRM needs authenticated API work, permissioning, and careful error handling. Every system it touches adds cost — and adds the most value.
- Data preparation. Your documents are messy: outdated PDFs, inconsistent formatting, knowledge that only lives in someone's head. Cleaning, structuring, chunking, and indexing that content for retrieval is real work and is the single most under-quoted line in most proposals.
- The accuracy bar. Getting a bot to "usually right" is fast. Getting it to "safe to put in front of paying customers" needs evaluation sets, prompt hardening, retrieval tuning, and graceful fallback when it does not know. This is routinely 30–40% of a serious build and it is exactly the part cheap quotes skip.
- Channels. Web widget only, or WhatsApp, Instagram, and in-app too? Each channel adds delivery, formatting, authentication, and testing overhead. WhatsApp in particular brings its own Business API approval and per-conversation costs.
- Language and tone. Multilingual, Hindi, or Hinglish handling adds prompt engineering and a lot of testing. Matching your brand voice is prompt work too.
- Compliance. If you are in finance, health, or handle personal data, add cost for data handling, logging, PII redaction, and audit trails.
A worked budget: mid-sized support bot
To make this concrete, here is a realistic one-time build for a RAG support chatbot answering from a company's help docs and order system, on web plus WhatsApp:
| Line item | Typical cost (INR) |
|---|---|
| Discovery, scoping, and success metrics | ₹25,000 – ₹50,000 |
| Data preparation and indexing | ₹40,000 – ₹1,00,000 |
| Retrieval + LLM integration | ₹80,000 – ₹2,00,000 |
| Order-system / API integration | ₹50,000 – ₹1,50,000 |
| Web widget + WhatsApp delivery | ₹40,000 – ₹1,00,000 |
| Evaluation, prompt hardening, QA | ₹60,000 – ₹1,50,000 |
| One-time total | ₹3,00,000 – ₹7,50,000 |
That spread is honest. The bottom of the range is a focused bot built by people who have done it before; the top is broader scope or a higher accuracy bar.
The costs nobody quotes upfront
The build price is a one-time number. These recur every month, and they decide whether the project is actually affordable:
- LLM API usage — usage-based, and hugely dependent on model choice and traffic. A low-traffic support bot might run ₹2,000–₹15,000/month; a high-volume consumer bot far more. This is the line that most surprises people at scale.
- Hosting and vector database — ₹1,500–₹20,000/month depending on scale and whether you self-host the retrieval layer or rent it.
- Maintenance and tuning — ₹10,000–₹50,000/month if you want someone watching quality, updating the knowledge base as your product changes, and correcting drift. A bot is not "done" at launch; an unmaintained bot slowly gets things wrong as reality moves on.
Budget for twelve months of running costs before you approve a build. A ₹3 lakh bot that costs ₹30,000/month to run is a ₹6.6 lakh decision in year one — and that is the number to compare against the salary of the support staff it replaces or augments.
Build vs buy vs no-code
- No-code SaaS platforms are the cheapest route to a flow bot and fine for simple deflection. They fall apart the moment you need real data access or custom actions, and their per-seat or per-message pricing quietly becomes your biggest line at volume.
- Buying a support tool with a built-in bot and customising it works well for standard customer support where you do not need deep integration.
- A custom build earns its cost when the bot touches your own systems, is a competitive differentiator, or runs at a volume where SaaS per-message pricing hurts — because integration and ownership are exactly what off-the-shelf tools cannot give you.
How to spend less without shipping a toy
- Scope one job first. A bot that deflects your top 20 support questions reliably beats a "does everything" bot that does nothing you can trust. Ship narrow, prove value, then expand.
- Pick the right model, not the biggest. Smaller, cheaper models handle most retrieval-grounded answers perfectly well; reserve frontier models for the genuinely hard 10% of queries and route intelligently. This one decision routinely cuts API bills by 5–10x.
- Own your data layer. Self-hosting the vector store and using open retrieval keeps you off per-message SaaS pricing and off a vendor's roadmap.
- Insist on an evaluation plan. If the proposal has no method to measure accuracy, you are paying to discover bugs in production, in front of customers. A test set is not optional for a bot that generates answers.
- Start with a pilot. A scoped 2–4 week pilot on your real data tells you more about feasibility and cost than any pitch deck.
How we approach it
At Navtechy we build RAG and agentic assistants the same way we build the rest of our products — Restrofi, InvoiceAI, and LeadsBuck — scoped to one clear job, wired into real data, and measured against an evaluation set before anything goes live. That discipline is the difference between a demo that impresses in a meeting and a bot your customers actually trust with their questions. If you want a straight, itemised estimate for your use case rather than a range off a homepage, tell us what the bot needs to do and what it needs to read, and we will cost it honestly.