---
title: AI Agent & Voice Agent Developer | India
description: AI agent development: chat agents from ₹75,000, voice agents from ₹1,50,000 plus ₹2–₹12 per minute. Built on your systems, with the running cost quoted upfront.
source: https://samuelmarndi.com/services/ai-agents/
site: https://samuelmarndi.com
author: Samuel Marndi
contact: +91 82803 20550 · samuelmarandi6@gmail.com
updated: 2026-08-31
---


# AI agent development — chat agents, voice agents, and the running costs nobody quotes

Agents that answer customers, act inside your systems and take phone calls — built on your data, with the monthly running cost written into the proposal instead of discovered in month three.

## Short answer

**An AI chat agent in India typically costs ₹75,000–₹4,00,000 to build, and an AI voice agent ₹1,50,000–₹8,00,000, with voice adding ₹2–₹12 per minute of conversation on top.** A retrieval agent that answers from your own documents sits at ₹6,00,000–₹15,00,000, and a multi-agent system that coordinates several workflows runs from ₹8,00,000 upwards.

**The number that decides whether an agent is worth building is the running cost, not the build cost.** Model tokens, speech-to-text, text-to-speech, telephony and hosting typically add ₹15,000–₹1,50,000 a month depending on volume. Any proposal that quotes a build price without a monthly figure attached is quoting half the project.

## Key facts

| Item | Detail |
| --- | --- |
| Chat agent build | ₹75,000 – ₹4,00,000 depending on integrations |
| Voice agent build | ₹1,50,000 – ₹8,00,000, plus ₹2–₹12 per minute |
| Retrieval agent on your documents | ₹6,00,000 – ₹15,00,000 |
| Multi-agent workflow system | from ₹8,00,000 |
| Running cost | ₹15,000 – ₹1,50,000 / month by volume — quoted before the build |
| Typical build time | 3–5 weeks chat · 5–10 weeks voice |
| Channels | Website, WhatsApp, phone, in-app, Slack, email |
| Delivery | Remote, worldwide. Based in Bengaluru. |

## What is delivered

- A defined scope: the exact tasks the agent handles and the ones it must refuse
- Integration with your real systems — CRM, calendar, ticketing, database, payments
- Retrieval over your own documents where answers must come from your content
- Guardrails: refusal rules, hand-off triggers, and no answering outside scope
- Human escalation on every path, including timeouts and errors
- A running-cost model with your projected volumes, before the build starts
- Evaluation set — real questions with correct answers, run against every change
- Every account, key, prompt and dataset in your name, exportable

## What an agent build costs, and where yours lands

These are Indian market ranges for 2026. The spread inside each band is driven almost entirely by integration count — an agent that answers from a document set is a fraction of the work of one that books appointments, checks stock and raises tickets across three systems that were never designed to be called by a machine.

**AI agent build cost in India, 2026**

| Item | Range | Notes |
| --- | --- | --- |
| Scripted FAQ bot | ₹30k – ₹75k | Fixed answers, no model reasoning |
| Chat agent | ₹75k – ₹4L | Model-driven, 1–3 integrations |
| Voice agent | ₹1.5L – ₹8L | Telephony, speech in and out |
| Retrieval agent | ₹6L – ₹15L | Answers from your own documents |
| Multi-agent system | ₹8L – ₹40L | Several coordinated workflows |

*Build cost only. Running costs are separate and shown below.*

## The running cost, which is the one that decides it

Build cost is one payment. Running cost is every month for as long as the agent exists, and it is the figure that determines whether the thing pays for itself. It is also the figure most proposals omit.

A voice agent is the clearest case. At ₹2–₹12 per minute all-in, a support line handling 2,000 calls a month at four minutes each costs between ₹16,000 and ₹96,000 a month to run — a six-fold spread that depends on choices made during the build. English-only is cheaper than Hinglish code-switching, which is cheaper than Tamil or Bengali. Reserved concurrency costs 15–30% more than best-effort. Committed volume cuts the rate substantially against pay-as-you-go.

I model this with your actual projected volumes before the build starts, and put the monthly figure in the proposal next to the build figure. If the arithmetic says the agent will not pay for itself, that is worth knowing in week one rather than month six — and I would rather tell you then.

| Cost layer | Chat agent | Voice agent | What moves it |
| --- | --- | --- | --- |
| Model tokens | ₹3,000 – ₹40,000 / month | ₹5,000 – ₹60,000 / month | Conversation volume and length, model choice |
| Speech to text | — | ₹0.30 – ₹1.50 / minute | Language, accent handling, accuracy tier |
| Text to speech | — | ₹0.50 – ₹3 / minute | Voice quality and whether it is a custom voice |
| Telephony | — | ₹0.40 – ₹1.20 / minute | Inbound vs outbound, carrier, volume |
| Hosting & vector store | ₹2,000 – ₹15,000 / month | ₹4,000 – ₹25,000 / month | Document volume, uptime requirement |
| Monitoring & tuning | ₹5,000 – ₹30,000 / month | ₹10,000 – ₹50,000 / month | How much you want it to keep improving |

## Which kind of agent do you actually need?

**Choosing between an agent, a chatbot and plain automation**

Most enquiries that arrive asking for "an AI agent" turn out to need one of the cheaper things below it. The honest answer is often the smaller build, and it is the one I will recommend.

- **If the answers never change and there are fewer than about 40 of them** → [A scripted FAQ bot](https://samuelmarndi.com/services/ai-consulting/) — ₹30,000–₹75,000, no model bills, nothing to hallucinate
- **If answers live in documents you already have** → [A retrieval agent](#pricing) — Answers cite your own content instead of inventing it
- **If the agent must do things — book, check, update, raise** → [A chat agent with tool access](#pricing) — The integrations are the work, not the conversation
- **If your customers phone rather than type** → [A voice agent](#pricing) — Adds per-minute cost; model the volume before building
- **If the task is the same every time with no judgement** → [Plain automation, not AI](https://samuelmarndi.com/services/custom-software/) — Cheaper, faster, deterministic, and it cannot hallucinate

## Why demos succeed and deployments fail

Almost every agent that disappoints does so for one of four reasons, and all four are decided at design time rather than discovered at launch.

The gap between a convincing demo and a reliable deployment is almost entirely guardrails, evaluation and hand-off — the unglamorous two-thirds of the work. A build quoted purely on the conversational part is quoting the demo.

- **No defined refusal set.** An agent with no explicit list of things it must not answer will answer them, confidently and wrongly. The refusal rules are written before the prompts.
- **No evaluation set.** Without a fixed set of real questions with known-correct answers, "we improved it" is an opinion. Every change is run against the set before it ships.
- **Retrieval used as decoration.** Pointing a model at your documents does not stop it inventing — it has to be constrained to answer from what it retrieved, and to say it does not know when nothing matches.
- **No human escape hatch.** Every path needs a route to a person, including the timeout and error paths. A customer trapped in a loop with a bot is worse for you than no bot.
- **Silent failure.** Agents degrade quietly when a model version changes or a document set drifts. Logging, alerting on refusal rates and a monthly review are part of the build, not an upsell.

## What you own at the end

Agent projects are unusually easy to get locked into, because the value sits in prompts, evaluation data and platform configuration that a vendor can simply keep. Same rule here as everywhere else on this site: it is yours.

Model provider accounts, telephony accounts and vector store are created in your business's name with you holding billing and the master credentials — so you can see exactly what the agent costs to run, and switch provider without asking me. Prompts, tool definitions, retrieval configuration and the evaluation set are handed over as files in your repository. If you replace me, the next developer starts from a working system rather than from archaeology.

## Pricing

Indicative Indian market ranges for 2026, published so you can sanity-check any agent quote — including mine. Your figure follows a scoping call, because integration count drives the price far more than conversation design does.

| Item | Indicative range | Notes |
| --- | --- | --- |
| Scripted FAQ bot | ₹30,000 – ₹75,000 | Fixed answers, no model reasoning, no token bill |
| Chat agent, 1–3 integrations | ₹75,000 – ₹4,00,000 | Website, WhatsApp or in-app; CRM and calendar typical |
| Voice agent | ₹1,50,000 – ₹8,00,000 | Telephony, speech in and out, call routing and hand-off |
| Retrieval agent on your documents | ₹6,00,000 – ₹15,00,000 | Ingestion, chunking, evaluation and citation of sources |
| Multi-agent workflow system | from ₹8,00,000 | Several coordinated agents across systems |
| Running cost — chat | ₹15,000 – ₹60,000 / month | Tokens, hosting, vector store, monitoring |
| Running cost — voice | ₹2 – ₹12 per minute | Language mix, concurrency and committed volume decide it |
| Evaluation & tuning retainer | ₹20,000 – ₹60,000 / month | Optional; keeps accuracy from drifting |

*Voice pricing moves fast — model and speech costs have fallen repeatedly — so the per-minute band is re-checked at quoting time rather than carried forward from an old proposal. Running costs are modelled on your projected volumes before the build is agreed.*

## How the work runs

1. **Scope and cost model** — Define exactly what the agent handles and refuses, then model the monthly running cost against your projected volumes — before anything is built.
2. **Build with an evaluation set** — Real questions with known-correct answers, assembled first, so every change can be measured rather than argued about.
3. **Guardrails and hand-off** — Refusal rules, escalation triggers and a route to a human on every path, including timeouts and errors.
4. **Launch, monitor, tune** — Logging and alerting on refusal and hand-off rates, with a monthly review while volumes settle.

## Frequently asked questions

### How much does it cost to build an AI agent in India?

A chat agent typically costs ₹75,000–₹4,00,000 depending on how many systems it has to integrate with, and a voice agent ₹1,50,000–₹8,00,000. A retrieval agent answering from your own documents runs ₹6,00,000–₹15,00,000, and a multi-agent system from ₹8,00,000. Indian delivery is commonly 60–70% below US agency rates for the same architecture.

### What does an AI voice agent cost to run?

₹2–₹12 per minute all-in, covering speech-to-text, text-to-speech, model tokens and telephony. A line handling 2,000 calls a month at four minutes each therefore costs roughly ₹16,000–₹96,000 monthly. English-only is cheapest; Hinglish code-switching costs more; regional languages more again. Committed volume cuts the rate meaningfully against pay-as-you-go.

### What is the difference between a chatbot and an AI agent?

A chatbot answers. An agent acts — it calls your systems to book an appointment, check stock, raise a ticket or update a record, and decides which of those to do. The conversation is the easy half; the integrations and the guardrails around them are the actual build.

### Will the agent make things up?

Not if it is built properly, and that is a design question rather than a model question. Answers are constrained to retrieved content, an explicit refusal set defines what it must not attempt, and it is required to say it does not know when nothing matches. An agent with no refusal rules will confidently answer things it should not — that is the default behaviour, not a defect you discover later.

### Can it work on WhatsApp?

Yes, and in India that is usually the right channel — most customers are already there. WhatsApp Business API, website chat, in-app, Slack, email and phone can all run off the same agent and the same integrations, so adding a channel later is a small increment rather than a second build.

### Can the agent answer from our own documents?

Yes — that is a retrieval agent. Your documents are ingested and indexed, and answers are generated from what is retrieved, with sources cited so a human can check them. Cost is driven by document volume, how messy the source material is, and how precise the answers have to be.

### How long does an agent take to build?

Three to five weeks for a chat agent with a couple of integrations, five to ten weeks for a voice agent. Most of that is not the conversation — it is integration work, the evaluation set, and the guardrail and hand-off design that separates a demo from something you can leave running.

### What happens when the agent cannot help?

It hands off to a person, on every path including timeouts and errors, with the conversation history passed across so the customer does not repeat themselves. This is designed first, not bolted on. A customer stuck in a loop with a bot damages you more than having no bot at all.

### Do we own the agent?

Yes. Model provider, telephony and vector store accounts are in your business's name with you holding billing and master credentials, so you can see the real running cost and change provider without me. Prompts, tool definitions, retrieval configuration and the evaluation set are delivered as files in your repository.

### Which models do you build on?

Whichever fits the task and the budget — commonly Claude, GPT and Gemini, with smaller or open models where the work is simple enough not to justify a frontier model. Model choice is a cost decision as much as a quality one, and it is reviewed at build time because prices and capabilities move every few months.

### Is an AI agent worth it for a small business?

Sometimes, and often not. If your volume is twenty enquiries a day, the running cost and the build rarely beat a person and a good FAQ page. It becomes worth it at genuine volume, out-of-hours demand, or repetitive lookups that block staff from better work. I model the arithmetic before the build and will tell you when the answer is no.

### Do you work with businesses outside Bengaluru?

Yes. Agent work is delivered remotely worldwide from Bengaluru — there is nothing physical to install. Scoping and reviews run over call, and everything is handed over in your own accounts and repository.

## Related services

- [AI consulting & chatbots](https://samuelmarndi.com/services/ai-consulting/) — Support chatbots, document AI, automation
- [Custom software](https://samuelmarndi.com/services/custom-software/) — Internal tools, billing, inventory, dashboards
- [App development](https://samuelmarndi.com/services/app-development/) — iOS & Android apps in Flutter or React Native
- [Website development](https://samuelmarndi.com/services/website-development/) — SEO-first sites, e-commerce and web apps


---

*Published by [Samuel Marndi](https://samuelmarndi.com), independent IT consultant, Bengaluru, India.
Source page: https://samuelmarndi.com/services/ai-agents/
Enquiries: +91 82803 20550 (phone and WhatsApp) · samuelmarandi6@gmail.com · Mon–Sat, 9am–7pm IST
Quotes are written and itemised, issued within 24 hours.*