Nearly every AI chatbot RFP we get in India starts with the same question: "What will this cost us for the whole year, not just the initial build?" Many founders have learned the hard way after vendors quote a low setup fee but end up charging four times more over the year for API tokens, message fees, and support retainers.
This post gives you the real numbers for 2026. We break down costs by chatbot type, show where ongoing expenses come in, and highlight the places where invoices often catch people off guard.
We build chatbots for our own SaaS products (VakeelSaathi, Dcomply, BookMySMS) and for clients every month. The numbers below reflect our real experience, not just what we wish the market looked like.
The five chatbot types and what each one costs
There isn’t one fixed price for an AI chatbot. There are five main types, and the cost difference between them can be as much as 40 times. Figure out which type you need first, and the budget will make more sense.
1. Rule-based / decision-tree bot (WhatsApp or website)
These don’t use any LLMs. They have fixed menus and simple if-this-then-that logic. They work well for things like order status, delivery tracking, appointment booking, and standard FAQs. You get 60 to 70 % of the value of a smart bot for just 10 % of the cost.
One-time build: Rs 50,000 to Rs 1.5 lakh
Monthly ops: Rs 3,000 to Rs 8,000 (hosting + maintenance)
Channel cost: WhatsApp session fees Rs 0.35 to Rs 0.85 per conversation; SMS Rs 0.15 to Rs 0.25
These are best for D2C brands, clinics, coaching institutes, and MSMEs where most user questions fit into a few templates. If you’re new to chatbots, this is usually the best place to start.
2. LLM-powered FAQ bot (no custom knowledge)
ChatGPT-style bot trained on generic knowledge, deployed on your website. Answers general questions in your brand voice but does not know your product or your data.
One-time build: Rs 1 lakh to Rs 2.5 lakh
Monthly ops: Rs 5,000 to Rs 15,000
LLM tokens: Rs 3,000 to Rs 20,000 per month depending on traffic (GPT-4.1-mini or Claude Haiku)
Honest take: this shape is dying. If you are going to spend on an LLM anyway, spend the extra Rs 1 to 2 lakh and get a proper RAG bot that actually knows your product.
3. RAG chatbot (LLM on your data)
This is what most Indian SaaS and enterprise buyers actually want when they say "AI chatbot". The bot reads your product docs, past tickets, policies, price lists or knowledge base and answers with paragraph-level citations. Handles 60 to 80 percent of L1 support tickets in production.
One-time build: Rs 3 lakh to Rs 8 lakh (single language, single knowledge base)
Complex builds: Rs 8 lakh to Rs 15 lakh (multi-lingual, multi-source, human handoff, analytics dashboard)
Monthly ops: Rs 25,000 to Rs 80,000
Monthly ops break down roughly as: LLM tokens Rs 10 to 40k, vector DB Rs 3 to 15k, embedding refreshes Rs 2 to 8k, monitoring and eval runs Rs 5 to 10k, human triage of low-confidence answers Rs 5 to 20k.
4. Multi-channel AI chatbot (web + WhatsApp + voice)
Same RAG brain, but reachable via WhatsApp Business API, IVR voice, Slack, or in-app chat. Every channel adds infrastructure, compliance, and per-message cost.
One-time build: Rs 6 lakh to Rs 15 lakh
Monthly ops: Rs 40,000 to Rs 1.5 lakh
Channel pass-through: WhatsApp Rs 0.35 to 1.35 per conversation (utility vs marketing), voice IVR Rs 0.60 to Rs 1.20 per minute, SMS Rs 0.15 to 0.25
Voice is the expensive one. Real-time speech-to-text plus LLM plus text-to-speech costs 3 to 5x a text conversation. Budget for it separately, not as a rounding error.
5. Agentic chatbot (takes actions, not just answers)
The bot does not just answer questions; it does things. Books the appointment in your calendar, updates the CRM record, applies the Stripe refund, opens the Jira ticket, sends the follow-up email. This is the shape that pays for itself the fastest, but also the one most vendors quote badly.
Pilot (one agent, one workflow): Rs 1.5 lakh to Rs 3 lakh across 3 to 4 weeks
Full agent suite (3 to 5 workflows): Rs 5 lakh to Rs 15 lakh across 6 to 10 weeks
Monthly ops: Rs 40,000 to Rs 2.5 lakh depending on volume + tool integrations
We wrote a longer piece on the agent-vs-chatbot decision if you want to go deeper. Short version: pick agents when the task needs 3+ steps across 2+ tools with a real decision in between.
Where the invoice usually surprises people
Six line items that founders underestimate almost every time.
LLM API costs at scale
An RAG bot serving 500 conversations a day with an average of 8 turns each and 2,000-token context per turn burns through roughly 2.4 crore tokens a month. On GPT-4.1, that is roughly Rs 12,000 to 30,000, depending on prompt caching. On Claude Sonnet 4.6 similar band. On GPT-4.1-mini or Haiku, it drops to Rs 3,000 to 8,000. Model choice per node in the graph matters more than the global model pick.
Vector database
Pinecone starts at USD 70 per month for the serverless tier and scales with namespace, throughput and index size. Weaviate Cloud sits in a similar range. pgvector on your existing Postgres is functionally free up to a few million vectors, which is why we default to it for anything under 5 million chunks.
Embeddings and re-indexing
Every time your knowledge base updates, chunks get re-embedded. OpenAI text-embedding-3-small is USD 0.02 per million tokens. Cohere embed-multilingual-v3 sits around USD 0.10 per million. A weekly full re-index of a 50k-document base is roughly Rs 300 to Rs 1,500 depending on the model. Small numbers individually, but they add up over the year.
Rerankers
Cohere rerank at scale is USD 2 per 1,000 searches. For a busy chatbot doing 15,000 queries a day, that is roughly Rs 75,000 a month just for reranking. If you skip the reranker, your answer quality drops noticeably. Do not skip it; budget for it.
Human review and low-confidence handoff
Every good RAG bot has a confidence threshold below which it says "I do not know" and routes to a human. Whoever answers that human queue is a cost. If it is a CSM already on payroll, folded into their week. If it is a dedicated support person, Rs 25,000 to Rs 60,000 per month depending on shift coverage. This is often ignored in pilot budgets.
Evals and drift monitoring
Nightly runs on a curated 200-question eval set catch model drift and retrieval regressions before customers do. Adds roughly Rs 3,000 to Rs 10,000 in tokens per month. Skipping this is the single fastest way to have a chatbot that quietly gets worse for two months before anyone notices.
What we actually see for a Rs 8 lakh build
To put real numbers behind the bands, here is a typical mid-market build we shipped last quarter. Single language, single knowledge base of 12,000 pages, deployed on client website plus WhatsApp Business API, with human handoff to a two-person support team.
Build cost: Rs 7.8 lakh across 9 weeks (discovery, ingest pipeline, retrieval, reranking, UI, WhatsApp connector, dashboard, evals, handover)
Monthly ops for the first year: Rs 62,000 average (Rs 18k tokens, Rs 8k vector DB, Rs 6k reranker, Rs 4k monitoring, Rs 26k retainer covering policy tuning, eval upkeep, model swap decisions, minor feature adds)
WhatsApp pass-through: Rs 32,000 average on 45,000 utility conversations per month
Total year one: Rs 7.8 lakh + Rs 11.3 lakh = Rs 19.1 lakh. They deflected roughly 61 percent of L1 tickets, freed 1.4 FTE of CSM time, and closed 8 percent more inbound sales conversations via the WhatsApp funnel. Payback was inside 7 months.
What to ask vendors before you sign
Five questions that separate the honest quote from the surprise invoice:
What is the model choice per node in the graph, and can we swap any of them next quarter without rebuilding? (Answer should mention LiteLLM or similar abstraction, not a hardcoded provider.)
What happens to our data? (Answer must specify zero-data-retention endpoints for enterprise vendors, or self-hosted Llama/Mistral if you are in BFSI or regulated.)
What does the monthly retainer actually cover? Policy tuning? Eval upkeep? Feature adds? Or just uptime? Get this in the SOW, not the sales deck.
What is the human handoff design and who staffs the low-confidence queue? If the vendor has no answer, they are quietly assuming your team does it for free.
Show us three past chatbots you have run in production for more than 12 months and their eval scores. If they cannot, they build pilots, not systems.
Where to start if you are budgeting today
Three sensible entry points depending on team size:
Under 50 employees: Rule-based WhatsApp bot for order or appointment flows. Rs 1 lakh build, Rs 5k/month. Get 70 percent of value at 5 percent of the cost.
50 to 300 employees: RAG bot on your product docs or past tickets. Rs 5 to 8 lakh build, Rs 40 to 60k/month ops. Deflects L1 tickets, pays for itself in 6 to 9 months.
300+ employees or regulated (BFSI, healthcare, legal): Multi-channel agentic build with proper compliance layer, self-hosted option, and dedicated SOC 2 or DPDP artifacts. Rs 12 to 25 lakh build, Rs 1.2 to 2.5 lakh/month ops. Longer payback but touches the whole customer journey.
If any of these bands make sense for your team, we run a paid 2-week discovery sprint that ends with a working prototype and a firm build quote. The full guide includes a tier-by-tier cost breakdown, orsend us your requirements, and we'll get back to you within 48 hours.