guide · comparison · costs
A chatbot, an AI agent, or hiring someone: what an Indian business actually gets
If customers keep asking your business the same questions, you have three options. Here is what each one really does, where each one fails, and what an AI agent costs in rupees, from a company that publishes its own bill.
A shop gets the same eight questions every day. What are your timings. Do you deliver to my area. How much is this. Can I return it. Where is my order.
You have three ways to stop answering them yourself. They are genuinely different things, and the difference is not the one the marketing suggests.
A rule-based chatbot
The kind with buttons. You draw a flow: if they tap "Delivery", show this message; if they tap "Returns", show that one.
What it is good at. It is predictable and cheap. For a small number of exact questions with exact answers — timings, an address, a price list — a button flow is honestly fine, and sometimes better than AI, because it cannot be wrong.
Where it fails. Real customers do not tap buttons, they type. "Bhai parcel 3 din se nahi aaya kya karu" does not match any button. A flow either sends them in a loop or dumps them to a human, and the customer has now had a worse experience than if you had said nothing.
Choose it if your questions are few, fixed, and you would rather a limited answer than a wrong one.
An AI agent
Reads the question, looks up your own material, writes an answer in the customer's own words and language.
What it is good at. It handles the question you did not anticipate. It reads "parcel 3 din se nahi aaya" and answers in Hinglish from your delivery policy. It works from your price list, your FAQ, your website, and you change its answers by changing that material rather than redrawing a flow.
Where it fails, and this is the part worth reading. It can be confidently wrong. If your knowledge says something out of date, it will say it with total conviction. If you let it answer without supervision, it can promise things you cannot deliver. This is a real risk and anyone selling you one should say so.
The honest mitigation is not better AI, it is approvals. Decide which actions need a person: let it answer delivery questions freely, make it ask you before promising a refund. Our agents show you the draft, you approve or edit it, and the edit is kept so the agent can be corrected.
Choose it if customers write to you in sentences, in more than one language, and your answers live in documents that change.
Hiring someone
A person on WhatsApp and email, answering.
What they are good at. Everything the other two are not. Judgement, an angry customer, a special case, a decision about money. A person can say "I will sort this out for you" and mean it.
Where it fails. Cost and hours. A support salary in India runs somewhere between about ₹15,000 and ₹30,000 a month depending on the city and experience, and that buys you one person, in one shift, in the languages they speak. Nobody answers at 11pm, and the question at 11pm is often the one from a customer about to buy.
Choose it if your volume justifies it, and — this is the real answer — choose it as well, not instead. The person should be spending their day on the hard third of your messages, not on "what are your timings".
The costs, honestly
We can only give you real numbers for ours, so here they are.
A typical reply from a Smart agent on Dhanur AI costs about ₹2.17. That's our published price on the pricing page, worked out from a typical reply's length at the model's price; every task then shows its exact cost, measured from our own ledger. A hundred customer questions a day is about ₹217, or roughly ₹6,500 a month.
We are not going to quote a competitor's prices at you, because their pricing changes and we would rather not be wrong about someone else's business in public. Check them yourself. What we will say is what to compare on: do you pay per conversation or per reply, is it in rupees with a GST invoice, and can you see what a single answer cost after it happens? If the answer to the last one is no, you cannot manage what you are spending.
And what you should hold against us: we charge for what the AI does, so a badly-set-up agent that searches your knowledge three times per answer costs more than a well-set-up one. We show you that per task rather than hiding it in a monthly fee, which we think is the right way round, but it does mean the bill varies.
What we would actually tell a shop owner
Start with the questions that have answers. Put your real material in — your price list, your delivery policy, your website — and let an agent handle those, with approvals on anything about money. Keep your person for the rest.
Do not start with WhatsApp, however much you want to, unless the tool can genuinely do it today. Ours cannot yet: the code is built and Meta has verified our business, but we do not have a production number, so we mark it Next on our status page rather than sell you something that is not there.
Start on your website or a shared chat link, see whether the answers are good, and expand from there. Every new workspace gets ₹25 of free credit, which is enough for about 11 replies on the Smart level, to try it on real questions before you decide anything.
Corrected on 2 October 2026: this post first said ₹25 was enough for thirty questions. At about ₹2.17 a reply on the Smart level, it is about 11.