Nobody publishes this number. Ask ten Indian development companies what an AI agent costs and you will get ten "depends on requirements" replies and a calendar link. So here is the honest version, including the parts that make us look expensive.
The short answer
For a business in India in 2026, a production AI agent lands in one of three bands.
| What you are buying | Typical range | Timeline |
|---|---|---|
| Single-channel agent, scripted flows, no system integration | Rs 60,000 – 1,50,000 | 2–3 weeks |
| Grounded agent with your data, one system integration (CRM or calendar) | Rs 1,50,000 – 4,00,000 | 3–6 weeks |
| Multi-channel with voice, several integrations, custom logic | Rs 4,00,000 – 12,00,000 | 6–14 weeks |
|---|
Plus running costs of roughly Rs 8,000 – 45,000 a month depending on conversation volume, which we break down below.
If someone quotes you Rs 15,000 for "an AI chatbot", they are selling you a decision tree with a language model bolted on the front. It will work for two weeks and then a customer will ask something the tree does not contain.
What actually drives the price
1. Whether the agent needs to know anything
A chatbot that answers from a fixed script is cheap. An agent that answers from your catalogue, your pricing, your policies and your inventory is a different piece of engineering. That requires retrieval — indexing your documents so the model answers only from them and says "let me check with the team" instead of inventing an answer.
This is the single biggest cost line, and it is the one worth paying for. An agent that invents a price or promises a delivery date you cannot meet costs more than it saves.
2. How many systems it has to touch
Reading from your catalogue is one thing. Writing into your CRM, checking a calendar for a genuinely free slot, raising a ticket, updating inventory — each of those is an integration, and each has its own auth, rate limits and failure modes.
A useful rule: every system the agent writes to roughly adds Rs 40,000 – 90,000 and a week of work, because you must also handle what happens when that system is down.
3. Voice or text
Voice costs more than text, and not by a small margin. You are adding speech recognition tuned for Indian accents and regional languages, text-to-speech that does not sound robotic, and — hardest of all — interruption handling. A caller who says "no wait, actually Thursday" mid-sentence must be understood.
Text agents on WhatsApp start around Rs 60,000. Voice agents realistically start around Rs 2,50,000.
4. Volume
Model and telephony costs scale with usage. At 1,000 conversations a month expect roughly Rs 8,000 – 15,000 in platform costs. At 20,000 conversations, Rs 60,000 – 1,20,000. Any proposal that does not mention running cost is hiding it.
The running costs nobody mentions
| Line item | Monthly, typical mid-size deployment |
|---|---|
| Model API (the AI itself) | Rs 4,000 – 25,000 |
| WhatsApp Business API conversations | Rs 3,000 – 18,000 |
| Voice minutes, if applicable | Rs 6,000 – 40,000 |
| Hosting and vector database | Rs 1,500 – 6,000 |
| Maintenance and prompt tuning | Rs 8,000 – 30,000 |
|---|
That last line is the one people skip, and it is the one that decides whether the agent still works in month six. Language changes, your products change, customers find new ways to phrase things. An unmaintained agent degrades.
What you can safely skip
- A custom-trained model. You almost certainly do not need to fine-tune anything. Retrieval over your own documents gets you 95% of the value at 5% of the cost. Anyone proposing to "train a model on your data" as the first step is adding a zero to the invoice for no benefit.
- A mobile app for the agent. Your customers are already on WhatsApp. Meeting them there costs nothing extra in adoption; an app costs everything.
- Multi-language on day one. Ship in the language 80% of your customers use, then add the second once you have real transcripts telling you which one it should be.
- A dashboard, at first. Everyone asks for analytics before they have any conversations to analyse. Ship the agent, run it a month, then build the dashboard around what you actually want to know.
How to tell whether a quote is honest
Ask these four questions. The answers tell you more than the price does.
- "What happens when the agent does not know?" A good answer describes an explicit fallback: it says so plainly and hands to a human with the conversation history attached. A bad answer is "it will figure it out."
- "Show me the agent being wrong." Anyone can demo a happy path. Ask to see it handle an out-of-scope question, an angry customer, and a half-typed message. If they cannot show that, they have not built it yet.
- "What is the monthly running cost at my volume?" If this is not in the proposal in writing, the project has an invisible bill attached.
- "Who owns the prompts and the data?" You should. If the agent's logic lives only inside a vendor's platform, you are renting your own customer conversations.
A realistic example
A real-estate business in Bhopal with about 400 enquiries a month. The agent qualifies budget, location and possession timeline on WhatsApp, checks live inventory, books site visits into the sales calendar, and escalates anything unusual to a human with full context.
- Build: Rs 2,80,000, six weeks
- Running: about Rs 19,000 a month
- Replaced: roughly 60% of the repetitive qualification work of two coordinators
- Caught: after-hours enquiries that previously went unanswered until morning
The payback argument was never "sack the coordinators." It was that enquiries arriving at 10pm on a Saturday were getting a reply at 10pm on a Saturday.
Where the money is actually wasted
Not on the build. It is wasted on building the wrong thing first. The most common pattern we see is a company automating its most visible process rather than its most expensive one, because the visible one is what leadership complains about in meetings.
Spend two weeks finding out which process actually costs you money before spending three months automating one that does not. That is the entire reason our engagements start with a fixed-price audit rather than a build.
FAQ
How much does a WhatsApp AI agent cost in India?
A production WhatsApp AI agent typically costs between Rs 60,000 and Rs 4,00,000 to build in India in 2026, depending on whether it answers from your own data and how many systems it integrates with. Running costs are usually Rs 8,000 to Rs 45,000 a month including model API, WhatsApp Business API conversation fees, hosting and maintenance.
How long does it take to deploy an AI agent?
A single-channel text agent takes two to three weeks. An agent grounded in your own documents with one system integration takes three to six weeks. Voice agents and multi-system deployments take six to fourteen weeks. WhatsApp Business API approval should be started in week one because it runs in parallel and is the most common cause of delay.
Is it cheaper to use a no-code chatbot platform?
Upfront, yes. No-code platforms cost a few thousand rupees a month and work well for simple FAQ deflection. They become expensive when you need the agent to read your live inventory, write into your CRM, or handle conversations that leave the script — at which point you are paying a subscription for something that still cannot do the job. The honest rule: use no-code to validate demand, then build properly once you know the agent earns its keep.
Do I need to fine-tune a model on my own data?
Almost certainly not. Retrieval-augmented generation, where the agent looks up answers in your documents at the moment it replies, delivers most of the benefit at a fraction of the cost and updates instantly when your catalogue changes. Fine-tuning is worth considering only for highly specialised language or very high volumes, and it should never be the first step.
What is the ongoing cost of running an AI agent?
Expect Rs 8,000 to Rs 45,000 a month for a mid-sized deployment: model API charges, WhatsApp or telephony fees, hosting, and prompt maintenance. Maintenance is the line most often omitted from proposals and the one that determines whether the agent still performs in month six.