WhatsApp AI Chatbots: 2026 Rules & Real Costs

Radiatus Engineering AI Published 31 Oct 2026 13 min read

Running a WhatsApp AI chatbot for business in India became clearer and dearer in 2026. Meta’s terms now keep general-purpose assistants such as ChatGPT off the Business Platform, but leave plenty of room for a clinic, shop or brand that uses AI to answer its own customers. The bigger change came on 1 October 2026: the replies your bot sends inside the 24-hour window are no longer free. This guide covers what Meta allows, recaps the rate card briefly, and works out what an AI conversation costs, in rupees, for a clinic and a D2C brand. Prices and rules are as of October 2026.

What Meta allows: business AI yes, general-purpose assistants no

In October 2025 Meta added an “AI Providers” clause to the WhatsApp Business Solution Terms. Providers of AI technology, such as large language models and general-purpose assistants, may not use the Business Platform when that technology is “the primary (rather than incidental or ancillary) functionality” of what they offer. It took effect on 15 January 2026 for providers already on the platform, and removed ChatGPT, Perplexity, Luzia and Poke (TechCrunch).

For most businesses nothing changed. Meta told TechCrunch that a travel company running a customer-service bot would not be barred. A clinic assistant that answers questions about timings and fees, books appointments and hands difficult cases to the front desk is incidental to the clinic: the product is the clinic, not the chatbot.

The grey zone is a business whose product is the conversation. If you are building an ask-anything health or legal assistant and WhatsApp is how people use it, you are the AI Provider the clause was written for. Take legal advice before you build.

One rule applies to every bot. Meta’s Business Messaging Policy requires automated experiences to offer a prompt, clear route to a human: transfer in the chat, a phone number, email or a support form.

Where the ban was lifted, and where it still applies

Italy’s competition authority ordered Meta to suspend the restriction in December 2025, and Brazil’s CADE imposed a preventive measure in January. On 15 January 2026, the day the ban took effect, Meta amended its terms so AI providers could still reach users with Italian (+39) or Brazilian (+55) numbers (9to5Mac).

In Europe, Meta offered rival chatbots paid access in March 2026. The European Commission judged that inadequate and, on 10 June 2026, ordered interim measures restoring the pre-October 2025 terms while its antitrust investigation continues (Engadget). OpenAI brought ChatGPT back to WhatsApp across the European Economic Area on 13 July 2026 (The Decoder).

India is less clear. As of 31 October 2026 we have found no action by the Competition Commission of India on this clause, and OpenAI’s July return named only the EEA, so the restriction appears to still apply to Indian (+91) numbers. One Indian outlet reported in July that some users could reach ChatGPT on WhatsApp again, while noting that neither company had confirmed it (Business Today). If your business model depends on the answer, read the current terms, not the news.

How WhatsApp charges in late 2026: a short recap

Meta has charged per delivered message since 1 July 2025. The rate depends on the message category and the country code of the recipient’s number, so a patient on a UK number is billed at the UK rate. Meta changes prices only on 1 January, 1 April, 1 July or 1 October (Meta pricing documentation).

The India figures come from Indian providers and press coverage of Meta’s 1 October 2026 rate card (Zymbler, Free Press Journal). We could not read them from Meta’s rate card files directly, so treat them as reported and check the INR rate card linked from Meta’s pricing page before you budget. They exclude GST; providers add 18%. Utility and authentication also have volume tiers at higher volumes.

What changed on 1 October 2026

• Service messages are charged again after being free since November 2024. Each business phone number gets 1,000 free delivered service messages a month. The allowance resets monthly and unused messages do not roll over.
• Utility templates inside an open window are charged. They had been free there since July 2025.
• No payment method, no replies. Meta stopped delivering service messages from 1 October for businesses with no payment method on file by 30 September 2026. If your bot went quiet that week, check billing before the code.
• INR billing. Eligible Indian customers must move all messaging accounts to INR billing by 31 December 2026 “to avoid disruptions”.

All four points are in Meta’s pricing documentation.

The 24-hour window and free entry points

When a customer messages or calls your number, a 24-hour customer service window opens, and each new message from them resets it. Inside the window you may send free-form messages; outside it, only approved templates. Your reminder does not open a window. Only the customer can.

When someone messages you through a click-to-WhatsApp ad and you reply within the customer service window, a free entry point window opens. Meta says it “may remain open for up to 7 days”; many older pages still say 72 hours. While it is open, Meta charges nothing for marketing, utility, authentication or service messages, but still charges for Meta Business Agent messages.

Two ways to run AI: Meta Business Agent or your own model

Small businesses on the WhatsApp Business app have a separate route: Business AI for small businesses in India, launched on 14 May 2026. It lives in the app, not the API, and Meta gives no price, so this guide does not cover it. On the Cloud API, which is where clinics with practice software and most D2C brands end up, there are two choices.

Meta Business Agent. Meta’s agent answers on your number. You give it business information, FAQs, files and your website, connect your APIs so it can take actions such as booking an appointment, and hand control to your app when needed (Meta Business Agent overview). Token billing at $2.00 per million tokens began on 1 August 2026 (Business Today). Meta estimates 20,000–25,000 tokens a message, about 4–5 US cents, and does not charge for cached tokens. Your business must be based in a country authorised for the agent, so confirm eligibility first.

Your own model. Your application receives the webhook, calls Claude, GPT or an open-weight model, and sends the answer back. Meta defines a service message as any non-template message not powered by its agent, whether a person or a third-party AI wrote it. So your AI’s replies use the 1,000 free messages, then cost the service rate, and your AI provider bills you separately for tokens.

Our view: at a few hundred AI replies a month, the agent’s price is easy to live with. At several thousand, its token charge becomes the largest line on the bill and your own model starts to pay for the work of building it. Where the crossover falls depends on your provider’s current prices, so do the sum.

Worked examples: what a month costs

Both examples are illustrative, not client work. They use the reported India rates above and assume every recipient has a +91 number.

A clinic: reminders plus an AI inbox

A dental or physiotherapy clinic on one number sends 2,000 appointment reminders a month as utility templates. Patients reply to confirm, reschedule or ask about parking, and the clinic’s own AI sends 1,800 replies inside open windows.

Your model’s cost goes on top:

monthly model cost = replies × (average input tokens × input price + average output tokens × output price)

Input tokens include the system prompt, the clinic information you retrieve and the conversation so far, which grows every turn unless you trim it. At an average of 4,000 input and 150 output tokens a reply, 1,800 replies use 7.2 million input and 270,000 output tokens. Multiply by your provider’s current rates.

Through Meta Business Agent the same 1,800 replies would use roughly 36–45 million tokens: $72–90 a month before tax, or ₹6,480–8,100 at an assumed ₹90 to the dollar. On the self-run route those replies cost ₹92 in Meta fees plus your token bill. The ₹230 for reminders is the same either way.

One detail catches clinics out. A patient messages at 9:40 am to ask about parking, and the reminder for her 5 pm appointment goes out at 10:00 am while her window is open. Until 30 September that utility template was free; now it costs ₹0.115. If your reminder engine checks for an open window first, it can send a free-form “See you at 5 pm today” instead. That is a service message, so it comes out of the free 1,000.

A D2C support bot

An online brand ships 6,000 orders a month with two utility templates each: confirmation, then dispatch with tracking. Its AI handles 1,500 support chats at four replies each. It sends 8,000 marketing templates for abandoned carts and restocks, and 800 shoppers arrive through click-to-WhatsApp ads.

Marketing is almost four-fifths of the pre-tax bill; the AI replies are ₹575. To cut cost, start with campaign targeting, not the bot. The engine choice matters more here, though. The same 6,000 replies through Meta Business Agent would be $240–300 a month, more than everything else on the bill combined, and its replies to the ad-driven chats would be charged too.

Designing a WhatsApp AI chatbot that costs less

Every delivered message is now a line item, so conversation design is a cost decision.

• Send one message, not three. A bot that answers in three bubbles pays three times. Write one reply and put choices in buttons or a list.
• Let buttons do the work. “Confirm”, “Reschedule” and “Talk to the front desk” resolve most reminder replies in one turn, with no model call.
• Keep deterministic steps out of the model. Checking a slot or an order status is an API call, not a reasoning task. Route it directly or through a workflow tool; our n8n self-hosting guide covers one way to run that glue.
• Prefer free-form replies inside an open window. They use the free 1,000; a utility template in the same window no longer saves anything.
• Keep reminders out of marketing. Add “20% off whitening this month” to a reminder and it becomes marketing, at about 7.5 times the price.
• Trim the prompt. Retrieve only relevant passages, cap history and cache the fixed part of the prompt where supported. Our guide to RAG in production covers retrieval and grounding.
• Route with a small model, answer with a larger one, and hand over to a person after a set number of unresolved turns.

Each number has its own free 1,000, so a clinic group with a number per branch gets one allowance per branch. Do not create numbers just to farm the free tier: it splits your inbox and saves little.

When an AI bot is the wrong answer

• Volume is low. Under a few hundred inbound messages a month, a receptionist with saved quick replies is cheaper and better.
• The questions need clinical judgement. The bot can collect details and book a slot. It should not interpret symptoms or reports, and anything urgent should go to a phone number at once.
• Your patients call rather than text. An AI voice agent has a different cost model and may fit better.
• The real problem is no-shows. Reminders with confirm and reschedule buttons need no model, and off-the-shelf tools beat a custom build. Our own Appointment Reminder App sends reminders through Meta’s Cloud API with approved templates, handles replies, keeps consent records, and adds an AI inbox built on the Claude API for the replies that need it.
• Your practice software already does it. Dental Software India and Physio Clinic Software include WhatsApp delivery. Start with what you have.

Compliance: opt-in, template categories and DPDP consent

Opt-in. Meta’s Business Messaging Policy lets you message people only if they gave you their number and opted in to hear from you, and you must honour opt-outs made on or off WhatsApp. Record when, where and how each person opted in.

Template categories. A utility template must be non-promotional and tied to the person’s own booking, order or account. Meta reviews approved templates on a recurring basis and can move one to a different category (template categorisation), multiplying its price. Keep offers in separate marketing templates.

DPDP consent. The Digital Personal Data Protection Rules, 2025 were notified in November 2025 with an 18-month phased timeline (PIB). They require a standalone, plain consent notice stating the specific purpose, and the Act requires withdrawal to be as easy as giving consent. For a WhatsApp assistant:

• Say up front that replies may come from an AI, what the chat is for and how to stop it.
• Collect only what the task needs. A booking bot needs a name and a slot, not a medical history.
• Know where conversations go: Cloud API provider, AI provider, your logs. Sign a data processing agreement with each processor and set a retention period.
• Make “stop” work across WhatsApp, SMS and calls at once.

Our DPDP Rules timeline and checklist covers the dates and full obligations.

Where to start

Price last month’s usage, by category and number, against the 1 October rates. Count how many inbound messages need a model rather than a button. Pilot, measure real token counts, then choose your engine. Fix opt-in records and template categories before you scale.

For a second opinion on the numbers or help with the build, Radiatus builds chatbots and virtual assistants, WhatsApp Business API integrations and AI agents that call your own systems, alongside wider generative AI work. The first consultation is free, and after a short discovery you get a written proposal with scope, milestones and price.

Frequently asked questions

Is a WhatsApp AI chatbot for business allowed in India?

Yes, if the AI serves your own customers. Meta’s January 2026 terms bar AI providers whose main product is a general-purpose assistant, but a clinic, shop or brand using AI for support, bookings or order updates is allowed. You still need opt-in and a clear route to a human.

What does a WhatsApp AI chatbot cost per month in India?

Meta’s share is often small. In our illustrative clinic example, 2,000 reminders and 1,800 AI replies come to ₹322 plus GST. Your AI provider’s token charges go on top. Meta Business Agent would instead cost about 4 to 5 US cents per reply. Build and maintenance costs depend on scope.

Are WhatsApp service messages still free?

Partly. Since 1 October 2026 each business phone number gets 1,000 free service messages a month, with no rollover. After that, Indian providers report ₹0.115 per message plus GST. Replies inside a free entry point window opened by a click-to-WhatsApp ad remain free.

What is Meta Business Agent pricing?

A global rate of 2 US dollars per million tokens, billed from 1 August 2026. Meta estimates 20,000 to 25,000 tokens per message, roughly 4 to 5 US cents. It is charged even inside free entry point windows and does not use the 1,000 free service messages.

Can a clinic’s WhatsApp bot answer medical questions?

It should not diagnose or interpret results. Use it for timings, fees, directions, bookings and reminders, let it collect details for the doctor, and send anything urgent or clinical to a person straight away.

Will WhatsApp prices change again?

Probably. Meta changes rates only on 1 January, 1 April, 1 July or 1 October, so check the rate card before each. Indian accounts must also move to INR billing by 31 December 2026.