In 2026, a WhatsApp AI agent costs roughly 4 to 10 lakh to build if it answers questions from your catalogue, documents and order data, and 10 to 25 lakh if it takes actions: booking a slot, tracking a shipment, raising a return, sending a payment link, updating the CRM. A menu bot with no AI in it is 50,000 to 2 lakh, or a monthly fee on a no-code platform. The AI figures sit on top of the ranges in our AI agent development cost guide: the same agent, plus one to two lakh of channel work that only WhatsApp needs.
The build is the smaller surprise. WhatsApp is the one channel where the platform itself sends you a bill for every message, and that bill changes on 1 October 2026, when Meta starts charging for the replies your agent sends inside the 24-hour window. Every quote written before that date is quoting yesterday's arithmetic. This guide has today's.
The quick-reference table
| What you are building | What it does | Build (INR) | Build (USD) | Timeline |
|---|---|---|---|---|
| Menu bot, no AI | Numbered options, fixed replies, a hand-off to a person. Built in a provider's flow builder or a few hundred lines of code | 50,000 to 2 lakh, or 1,000 to 10,000 rupees a month on a platform | $600 to $2,400 | 1 to 3 weeks |
| Answering agent | Understands free text in Hindi, English and Hinglish, answers from your catalogue, FAQs and policies, looks up an order by number, hands off everything else | 4 to 10 lakh | $4,800 to $12,000 | 4 to 6 weeks |
| Acting agent | Everything above, plus bounded actions: books a slot, raises a return, sends a payment link, updates the CRM, sends the template that follows, with limits and an audit log | 10 to 25 lakh | $12,000 to $30,000 | 6 to 10 weeks |
| Multi-flow or commerce agent | Several journeys (sales, support, collections) across many systems, catalogue and cart inside WhatsApp, payments, voice notes, approval gates | 25 lakh and up | $30,000 and up | 3 months and up |
The line that matters is between the second and third rows, and it is the same line as in every agent project. An agent that only answers is a retrieval problem. An agent that acts needs action limits, an audit trail and a hand-off that works, and the difference is spelled out in chatbot vs AI agent. On WhatsApp there is a second line, between the first row and the second: a menu bot is not an AI agent, and a good share of the "WhatsApp AI chatbot" quotes in the Indian market are pricing the first row while describing the second.
Three bills, not one
A WhatsApp agent runs up three separate bills, and a budget that shows one of them is a third of a budget.
- The build. Paid once, to whoever builds it. The tables and the worked estimate below.
- Meta's fees. Paid per message, to Meta, through your provider or directly through the Cloud API. Set by Meta, by country and by message category, and revised on a published schedule.
- The running cost. Model usage per conversation, hosting, monitoring, and the monthly fee of the provider you send through. The general arithmetic is in what an AI agent costs to run; the WhatsApp version below adds Meta's line to it.
Most quotes we see cover the first, gesture at the third and leave out the second entirely. From October the second is no longer small enough to leave out.
What Meta charges, and what changes on 1 October 2026
Meta charges per message, by category, at a rate set per country. The category is decided by the template you send, or by the fact that you are replying inside a conversation the customer started. At the time of writing, Meta's India rate card puts the categories roughly where the table below has them. Check the current card before you budget, because Meta revises it on a quarterly schedule and did so for India in January and April 2026.
| Message category | What it is | India rate, September 2026 | From 1 October 2026 |
|---|---|---|---|
| Marketing template | Offers, cart recovery, re-engagement, anything promotional, sent to customers who opted in | About 86 paise per message | Unchanged |
| Utility template | Order confirmation, shipping update, appointment reminder, payment reminder | About 12 to 15 paise outside the 24-hour window, free inside it | Charged inside the window too |
| Authentication template | One-time passcodes | About 12 to 15 paise | Unchanged |
| Service message | A free-form reply, from a person or from your AI agent, inside the 24 hours after the customer last messaged you | Free | Charged at the utility rate after the first 1,000 a month per phone number |
The last row is the one that changes the economics of an AI agent. Meta announced in July 2026 that from 1 October, every free-form reply your agent sends inside the 24-hour customer service window is billed at the same per-message rate as a utility template, with 1,000 free messages per business phone number each month and no volume tiers. Until now those replies were free, which is why so many WhatsApp agents were built to send four short messages where one would do.
Two things stay free. A conversation that starts from a click-to-WhatsApp ad, or from the call-to-action button on a Facebook or Instagram page, opens a 72-hour window in which every message of every type is free. And Meta's own Business Agent is billed on its own usage meter rather than per service message, which is Meta's way of nudging you toward it. We come back to that in the build-or-buy section.
The practical consequence: from October, a WhatsApp agent has a cost per reply, and the number of replies it takes to finish a conversation is a cost lever in a way it never was before.
Where the build money goes: a worked estimate
Tier tables hide the arithmetic, so here is an estimate for a concrete agent. Take a direct-to-consumer brand selling on its own store and on marketplaces, with 20,000 WhatsApp conversations a month: where is my order, how do I return this, is this available in a medium, I paid but the order still shows pending. The agent reads each message in Hindi, English or Hinglish, identifies the customer and the order, answers from the catalogue and the policies, tracks the shipment with the courier, raises a return or an exchange inside a rule set, sends a payment link for a pending order, and hands everything else to a person in the shared inbox with a summary attached. This is how we would break it down after discovery.
| Work package | Person-days | Share of budget |
|---|---|---|
| Discovery: reading real chat exports, listing the intents, agreeing what "finished" means for each | 5 | 6% |
| Meta setup: business verification, the WhatsApp Business Account, the number and display name, choosing a provider or the Cloud API, first template submissions | 4 | 5% |
| Collecting and labelling 300 real conversations with the correct outcome, and building the scorer | 8 | 9% |
| Prototype: retrieval over catalogue and policies, language handling, first prompts, scored against the set | 10 | 11% |
| Connectors: order system, courier tracking, returns, payment links, CRM | 15 | 17% |
| WhatsApp channel work: buttons and lists, images and PDFs, template flows for updates and cart recovery, the 24-hour window logic, opt-in records | 10 | 11% |
| Guardrails: action limits, approval gates, prompt-injection defences, personal data handling | 7 | 8% |
| Hand-off to a person inside the shared inbox, with the conversation summarised | 6 | 7% |
| Observability: a trace per conversation, an audit log, a dashboard showing cost per conversation with Meta's line in it | 6 | 7% |
| Testing: adversarial messages, regression against the set, load, template rejection paths | 7 | 8% |
| Staged rollout and a month of tuning, feeding new cases into the set | 10 | 11% |
| Total | 88 |
At a blended rate of 16,000 rupees per person-day, mid-band for an experienced Indian team, that is about 14 lakh. With two engineers on it, nine weeks. Both numbers sit in the third row of the table above, and the shape matches the 80-day support agent in our agent cost guide: the extra eight days are the Meta setup and the channel work, which is the whole of the WhatsApp premium.
Notice what is not the expensive part. The prototype, the piece with the model in it, is 10 of 88 days. Connectors, guardrails, testing and rollout are 39. A quote that skipped those would come in near 6 lakh for what looks like the same agent in a demo, and it would be the agent that approves a return on the wrong order in month two.
The monthly bill at 20,000 conversations
The same agent, live, in a month after October 2026. Assume four agent replies per conversation, 10,000 marketing templates a month for cart recovery and offers, and 15,000 utility templates for order updates.
| Item | Monthly amount |
|---|---|
| Service messages: 80,000 replies, less 1,000 free, at about 15 paise | About 12,000 rupees |
| Utility templates: 15,000 at 12 to 15 paise | 2,000 to 2,500 rupees |
| Marketing templates: 10,000 at about 86 paise | About 8,600 rupees |
| Provider fee, or nothing on the Cloud API directly | 0 to 10,000 rupees |
| Model usage: 80,000 calls at 2,000 input and 120 output tokens, 30 percent overhead, at an illustrative $3 and $15 per million | About $810, roughly 68,000 rupees |
| Hosting, search index, monitoring | 8,000 to 15,000 rupees |
| Total | Roughly 1 to 1.2 lakh a month, or 5 to 6 rupees a conversation |
Two things in that table should change how you build. First, the model line is the largest and the most movable: route the routine turns, which are most of them, to a small model and keep the capable one for the hard ones, and the line falls to around 30,000 rupees. Provider-side caching of the fixed part of the prompt takes it lower still. Second, Meta's line is now about a fifth of the bill and it scales with replies, not conversations. An agent tuned to close a conversation in two messages instead of four halves it.
For comparison, a support executive at 25,000 to 40,000 rupees a month handles somewhere between 1,500 and 3,000 WhatsApp conversations, which is 12 to 25 rupees each. The agent at 5 to 6 rupees, or 3 to 4 with the model split, is not the whole story, because the hard cases still reach a person and they take longer than the average did. Count the agent's value on the conversations it finishes, not the ones it touches.
The seven things that move the price
- Answering or acting. The largest step, as in every agent project. An agent that tells a customer where the order is sits in the second row. One that raises the return, moves money or changes the booking needs limits, gates and logs, and sits in the third.
- Languages and script. Hindi and English are table stakes. Hinglish typed in Latin script, regional languages, and customers who switch language mid-conversation each add examples to the test set and turns to the prototype. Budget a week per additional language for a proper evaluation, not a day.
- The number of systems. Every system the agent reads from or writes to is a connector, a permission model and a set of failure cases. An order system and a courier is a normal first agent. Add a CRM, a payment gateway, a loyalty platform and a marketplace and you have doubled the connector line.
- Commerce inside WhatsApp. A catalogue and a cart inside the chat, with a payment at the end, is a different build from a support agent: product data feeds, stock sync, price changes and the payment flow all have to hold. It moves the project into the fourth row.
- Provider or Cloud API. Going through a provider is faster to start and costs a monthly fee plus a markup on every message. Going direct to Meta's Cloud API costs a few more days of build and nothing per month beyond Meta. The crossover is a volume question, and the build-or-buy section has the arithmetic.
- Templates and marketing volume. Every outbound flow is a template that Meta has to approve, and every marketing message is the most expensive category by a factor of six. An agent that sends a lot of outbound needs the template library designed, submitted and priced.
- Consent and data. Marketing templates need a recorded opt-in. Customers share addresses, phone numbers and payment details in the chat, so retention, deletion and where the transcripts live all need decisions. Regulated categories, health and finance in particular, add a review outside engineering.
Build, buy or configure
You do not always need a custom agent, and a good agency will say so. Four options cover most cases.
| Option | What you pay | Fits when |
|---|---|---|
| A provider's no-code bot builder (WATI, Interakt, AiSensy, Gupshup and others) | 1,000 to 10,000 rupees a month, plus a markup of 20 to 85 paise on each message over Meta's rate, at the time of writing | Menu flows, broadcasts and a shared inbox, with no AI or a simple FAQ layer |
| A provider's AI add-on | A higher plan, sometimes priced per conversation | Answering from a help centre in one language, with no actions in your systems |
| Meta Business Agent | Billed by Meta on its own usage meter, separate from service messages | Answering inside Meta's surfaces with Meta's model, when the agent never needs to act in your systems |
| A custom agent, on the Cloud API or through a provider | 10 to 25 lakh to build, plus the monthly bill above | Actions across more than one system, your own data and rules, several languages, or volumes where the per-message markup costs more than owning it |
The arithmetic on the markup is worth doing. At a markup of 50 paise, 100,000 messages a month is 50,000 rupees a month to the provider, six lakh a year, for plumbing you could own. Below 20,000 messages a month the provider is cheaper and simpler and we would tell you so. Many of our builds use a provider's inbox for the people and a custom agent for the judgement, on the same number, which is usually the cheapest combination of all.
The costs that are not in the quote
| Cost | Typical size | When it surprises you |
|---|---|---|
| Business verification | Days to weeks of waiting rather than money | When the build is done and the number is not yet cleared to message at volume |
| Template rejections | A day or two per round | When the cart-recovery template is classed as marketing after it was priced as utility |
| Messaging limits and quality rating | Throughput rather than money | When a broadcast gets the number rate-limited and the agent's replies queue behind it |
| Opt-in collection | A flow on the website, at checkout and in the chat | When you find that marketing templates cannot go to customers who never opted in |
| Meta rate changes | A revised card each quarter | When the October change adds a line the annual budget did not have |
| Model retirements | 2 to 5 days of re-tuning, once or twice a year | When the provider retires the model you launched on and the prompts behave differently |
| Provider lock-in | A migration project | When the templates, the opt-ins and the number itself sit in the provider's account rather than yours |
| The people the agent hands off to | Training, and a harder case mix | Month two, when the easy conversations are gone and the average human case takes longer |
GST at 18 percent applies to the development invoice as it does to any software work in India, and it is claimable as input credit if you are registered, but the cash goes out first. Meta's fees are billed in rupees for Indian businesses since January 2026, with GST on top.
How to spend less without a worse agent
- Close in fewer messages. From October every reply has a price. An agent that answers the question, confirms the action and offers the next step in one message costs a quarter of one that does it in four, and reads better.
- Put the fixed answers in a menu or a template. Store hours, return policy, delivery areas: a list message with the answer attached costs one message and no model call. Save the model for the questions that need it.
- Route the plain turns to a small model. Classifying a message and fetching an order status does not need the model that handles a complaint. The split cuts the model line by half to two thirds.
- Start read-only with a person sending. Let the agent draft replies in the shared inbox for the first month. You collect examples and confidence data for free, and the acting version is cheaper to build on top of it.
- Use the 72-hour ad window. Conversations that start from a click-to-WhatsApp ad are free for three days, whatever you send. If you run those ads, the agent should be on that number.
- Start verification and templates on day one. Business verification, the number, the display name and the first templates all run on Meta's clock. A rejection in week seven costs a week; the same rejection in week one costs nothing.
- Own the account. The WhatsApp Business Account, the number, the templates and the opt-in records should sit in your Meta Business portfolio, with the provider granted access, not the other way round.
How to read a WhatsApp agent quote
Six checks, ten minutes.
- Ask which row of the table it is pricing. A "WhatsApp AI chatbot" at 40,000 rupees is a menu bot with a name. That can be the right purchase, but it should be sold as what it is.
- Ask for Meta's line as a separate number, at October rates. A quote that says "WhatsApp messages are free within 24 hours" was written before July and has not been updated since.
- Ask who owns the WhatsApp Business Account, the number and the templates. The answer should be you, in writing, from day one.
- Ask for the test set. If the quote does not include collecting real conversations and scoring against them, ask how anyone will know the agent works in Hindi as well as it does in the English demo.
- Ask what the agent may do without a person. The list should be short, explicit and limited by amount or type. If the answer is "anything the API allows", the guardrails are not in the price.
- Ask how a conversation reaches a person, and what the person sees. The hand-off is where WhatsApp agents live or die, because the customer is still in the same chat and can tell when nobody is there.
The longer checklist for choosing the vendor itself is in our guide to hiring an AI development company.
Questions people ask
How much does a WhatsApp chatbot cost in India?
A menu bot with fixed replies costs 50,000 to 2 lakh to build, or 1,000 to 10,000 rupees a month on a provider's platform. An AI agent that understands free text and answers from your own data costs 4 to 10 lakh; one that takes actions in your systems, with limits and an audit log, costs 10 to 25 lakh. Both AI figures include the one to two lakh of Meta setup and channel work that a WhatsApp deployment needs and a web chatbot does not.
How much does a WhatsApp AI agent cost per month?
For the agent in the estimate above, at 20,000 conversations a month, roughly 1 to 1.2 lakh a month after 1 October 2026: about 12,000 rupees in service messages, 10,000 to 11,000 in templates, up to 10,000 to a provider, around 68,000 in model usage at mid-tier prices, and 8,000 to 15,000 in hosting and monitoring. Routing routine turns to a small model brings the model line down to about 30,000, and the total to 3 to 4 rupees a conversation.
Is the WhatsApp Business API free?
The API itself has no licence fee, and you can use Meta's Cloud API directly with no monthly charge. What you pay for is messages: marketing, utility and authentication templates at per-message rates set by country, and from 1 October 2026 the free-form replies your agent sends inside the 24-hour window, after 1,000 free messages a month per number. Providers add their own platform fee and markup on top.
What changes on 1 October 2026?
Service messages, meaning free-form replies inside the 24-hour customer service window, stop being free. Meta bills them at the utility rate for the recipient's country, which in India is around 12 to 15 paise at the time of writing, after 1,000 free messages per business phone number per month. Utility templates sent inside the window are charged too. Conversations that start from a click-to-WhatsApp ad keep their 72-hour free window.
Can I build a WhatsApp AI agent with a no-code tool?
For a menu bot, yes, and for an answering agent over a clean help centre in one language, often yes. No-code stops working when the agent has to act in more than one system, when you need action limits and an audit log, when customers write in three languages, or when the provider's per-message markup at your volume costs more than owning the agent would. Those points arrive sooner on WhatsApp than on web chat, because WhatsApp volumes are higher and the messages are shorter.
How long does it take to build a WhatsApp AI agent?
Six to ten weeks for an acting agent: a week of discovery, two weeks of prototype scored against real conversations, then five to seven weeks of connectors, channel work, guardrails, testing and a staged rollout. An answering agent takes four to six weeks. Meta's business verification runs alongside and should start in week one, because it is the one step nobody can speed up.
Does it work in Hindi and regional languages?
Yes, if it is built and tested for them. The models handle Hindi, Hinglish and the major regional languages well. What fails is an agent tested only in English: the test set has to contain real messages in every language your customers use, and the retrieval has to find the right policy when the question is in Hindi and the policy is in English. That is evaluation work, and it is in the estimate above.
Getting a real number
Ranges end where your actual conversations begin. We price WhatsApp agent projects after a one-week discovery and a two-week prototype scored on your own chat history, so the number you approve is based on how the agent performs in your customers' languages, not on how it demos in ours. What that process looks like is on our AI agent development page, and if the agent also needs to answer calls, the voice side is on our voice AI page.