Every second conversation I have with a business owner now includes the words "AI agent". Usually it comes after a vendor demo: a bot that answers customer queries, chases payments or updates the CRM on its own. The demo looks magical. The quote looks big. And the owner wants to know one thing: is this real, or am I about to waste lakhs?
Almost everything you'll find online about AI agents is written by companies that sell them. This guide is written from your side of the table.
First, what is an AI agent, really?
Vendors use "agent" for almost anything with AI in it. Gartner calls this "agent washing" and estimates only about 130 of the thousands of vendors claiming agentic AI actually offer it. So before you compare quotes, know which of these three you're being sold:
| Compared on | Chatbot | Automation | AI agent |
|---|---|---|---|
| What it does | Answers questions from a script or a knowledge base | Runs fixed steps when something happens ("when an order comes in, send an invoice") | Works towards a goal: reads the situation, decides the next step, uses your tools, and checks the result |
| Handles the unexpected? | Barely | No, it breaks or skips | Yes, within limits you set |
| Takes actions in your systems? | Rarely | Yes, the same ones every time | Yes, and chooses which ones |
| Example | Website FAQ bot | Tally entry from a Zoho form | Reads a dealer's WhatsApp order, checks stock and credit limit in your ERP, books it, and flags exceptions to a person |
| Typical build | Days to weeks | Days to weeks | Weeks to months |
Chatbot
- What it does
- Answers questions from a script or a knowledge base
- Handles the unexpected?
- Barely
- Takes actions in your systems?
- Rarely
- Example
- Website FAQ bot
- Typical build
- Days to weeks
Automation
- What it does
- Runs fixed steps when something happens ("when an order comes in, send an invoice")
- Handles the unexpected?
- No, it breaks or skips
- Takes actions in your systems?
- Yes, the same ones every time
- Example
- Tally entry from a Zoho form
- Typical build
- Days to weeks
AI agent
- What it does
- Works towards a goal: reads the situation, decides the next step, uses your tools, and checks the result
- Handles the unexpected?
- Yes, within limits you set
- Takes actions in your systems?
- Yes, and chooses which ones
- Example
- Reads a dealer's WhatsApp order, checks stock and credit limit in your ERP, books it, and flags exceptions to a person
- Typical build
- Weeks to months
In practice, a lot of what businesses need is automation, not an agent. It's cheaper, more predictable and easier to fix. An agent earns its cost only when the work genuinely needs judgement, like reading messy messages, handling exceptions, or deciding between options.
Where an agent pays off, and where it doesn't
Two questions decide it: how often does the task happen, and what does one mistake cost you?
Build an agent where the task is frequent and a mistake is cheap
A two-by-two grid. Vertical axis: cost of one mistake, from low to high. Horizontal axis: how often the task happens, from rarely to hundreds of times a day.
- Rare tasks where a mistake is costly: keep it with people. Examples: negotiating a contract or price, handling a serious complaint, hiring and HR decisions.
- Frequent tasks where a mistake is costly: the agent drafts, a person approves. Examples: refunds and credit notes, quotes and pricing replies, payment reminders to key accounts.
- Rare tasks where a mistake is cheap: use an off-the-shelf AI tool. Examples: summarising meetings, drafting the odd email or policy, one-off internal questions.
- Frequent tasks where a mistake is cheap: build an agent. Examples: order-status and delivery queries, capturing dealer orders from WhatsApp, reading invoices into your ERP, qualifying website enquiries.
The sweet spot is the bottom right: work your team does dozens or hundreds of times a day, where an error is easy to spot and cheap to fix. Top right is still worth it, but the agent should prepare the work and a person should approve it. On the left side, the volume rarely justifies building anything custom.
A quick test before you spend anything: write down the task, how many times a week it happens, how long it takes your team each time, and what one mistake costs. If the hours saved in a year don't comfortably beat the build cost plus a year of running costs, it isn't time for an agent yet.
What an AI agent costs in India
There are two bills: building it, and running it every month. Most quotes only show the first.
1. The build
| What you're building | Typical build cost (India, 2026) | Timeline |
|---|---|---|
| An AI feature added to an existing app or system | ₹3–15 lakh | Weeks |
| A focused agent or AI copilot on a hosted model, using your documents | ₹10–21 lakh | 10–16 weeks |
| A full AI-native product (agents, retrieval, voice, testing) | ₹20–60 lakh | 5–9 months |
These are market ranges from published Indian estimates, not quotes. Simple automations built on tools like n8n or Make can cost far less. Custom-trained models cost far more.
2. The monthly running cost
Every time an agent reads or writes, you pay the AI model provider by the amount of text processed ("tokens"). This is the cost owners forget, and it grows with every customer you add.
A rough example: a WhatsApp support agent handling 500 conversations a day, about six messages each, on a mid-priced model. That's roughly 15,000 conversations a month and somewhere around ₹50,000–60,000 a month in model costs alone. A top-end model can cost about three times that. Add hosting, a vector database for your documents, WhatsApp message fees and monitoring.
Then add maintenance. Plan for 15–25% of the build cost every year: models change, prompts need tuning, and your business rules move.
The question to ask before signing: "What will this cost per month at today's volume, and at 3x our volume?" If the vendor can't answer with a number, they haven't done the maths. You'll find out on your cloud bill.
Build, buy or configure: three ways to get an agent
Buy it inside a tool you already use.
Your CRM, helpdesk, accounting or WhatsApp platform may already offer an AI agent as an add-on. It's the fastest and cheapest route, and the right one for standard tasks like answering FAQs or summarising tickets. The limit: it only knows what that tool knows, and you're tied to its pricing.
Configure it on a platform.
Tools like n8n, Make or the agent builders from the big AI providers let a developer wire an agent to your systems without writing everything from scratch. Good for internal workflows across two or three tools. Watch who owns the setup, and whether anyone on your side can change it.
Build it custom.
Worth it when the agent is your advantage: it touches your core operations, uses your proprietary data, or has to work in ways no tool supports. It costs the most, and needs someone accountable for it long after launch.
My rule of thumb: start with the first route that genuinely works, and only move down the list when you hit its limits. Most businesses I see jump straight to "build" because that's what the vendor in the room sells.
The risks owners underestimate
Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear business value or weak risk controls. In my experience, these are the four that catch owners out:
01Nobody defined "good".
If you can't say what a correct answer looks like, and how often the agent must get it right, you can't tell whether it's working. Agree on a measurable target before the build, like "resolves 60% of order-status queries with no human, with under 2% errors".
02It's confidently wrong.
AI models can make things up and sound sure about it. An agent quoting the wrong price, promising a delivery date you can't meet or approving the wrong refund is a real cost. Keep a person in the loop for anything involving money, commitments or complaints, at least until the numbers prove it's safe.
03Your data goes somewhere.
Customer messages, invoices and order history travel to the AI provider's servers. Under India's DPDP Act, you're responsible for that data, including what your vendors do with it. Ask where it's processed, whether it's used to train models, and how long it's kept. (I've written a separate guide on what the DPDP Act means for your systems.)
04Nobody owns it after launch.
Agents drift. Your products, prices and policies change; the model provider updates its models; edge cases pile up. Without someone reviewing conversations every week and adjusting the agent, quality quietly falls.
10 questions to ask an AI agent development company
Before you sign with anyone, send these. The answers tell you more than the demo.
- Is this an agent, an automation or a chatbot? Why does our problem need that one?
- What exactly will it do on its own, and what will it hand to a person?
- How will we measure if it's working? What accuracy do you commit to, and how is it tested before launch?
- Which AI model will it use, and can we switch models later without rebuilding?
- What will it cost per month to run at our current volume, and at 3x?
- Where is our data sent and processed? Is it used to train anyone's models? How long is it kept?
- What happens when it gets something wrong? Can we see every conversation and decision it made?
- What systems will it connect to, and what can it change in them? Can it be limited to read-only at first?
- Who maintains it after launch, what does that cost, and what's the response time when it breaks?
- Who owns the prompts, the code, the workflows and the data if we stop working together?
Be wary of a vendor who can't answer 3, 5 and 7 clearly. Those are where agent projects quietly fail.
Frequently asked questions
What is an AI agent for business?
Software that uses an AI model to work towards a goal on its own: it reads a situation, decides the next step, uses your business tools (like your CRM, ERP or WhatsApp), and checks the result. Unlike a chatbot, it takes actions. Unlike a simple automation, it can handle cases that don't follow a fixed script.
How much does it cost to build an AI agent in India?
A focused agent on a hosted AI model typically costs ₹10–21 lakh to build, and a complex AI-native product ₹20–60 lakh. Adding a simpler AI feature to an existing system can cost ₹3–15 lakh. Running costs are extra and grow with usage.
What does an AI agent cost to run every month?
It depends on volume and the AI model used. A support agent handling around 15,000 conversations a month on a mid-priced model might cost ₹50,000–60,000 a month in model fees, plus hosting and maintenance. Always ask for a monthly estimate at your current volume and at 3x.
Is an AI agent better than hiring a person?
For high-volume, repetitive work with clear rules, an agent can handle the bulk and let your team focus on exceptions. For judgement calls, relationships and anything where one mistake is expensive, people are still better. Most good setups combine both.
Can I use an AI agent on WhatsApp?
Yes. WhatsApp is one of the most common channels for agents in India, through the WhatsApp Business Platform. Factor in WhatsApp's own message fees and Meta's rules on what businesses can send.
Do AI agents comply with the DPDP Act?
The agent doesn't comply; your business does. If it processes customers' personal data, you're responsible for consent, security and how your AI vendor handles that data.