WhatsApp Business Calling API with an AI Agent: Setup, Costs and What Breaks

The WhatsApp Business Calling API, which Meta launched in 2025, lets a business place and receive voice calls inside a WhatsApp thread. A customer taps the call button on your verified WhatsApp number and the audio travels over WhatsApp’s own infrastructure — there is no separate phone line, no DID to provision and no PSTN routing in the path.

That single detail is what makes it interesting. A support call and the chat history now live in the same conversation, so an agent picking up already has the thread in front of them. The caller does not dial a number, wait through a menu, or repeat what they typed five minutes earlier.

• 535M+WhatsApp users in India
• 11Indian languages Adya answers in
• 250M+Indians who code-switch daily
• <400msLatency a natural call needs

India is WhatsApp’s largest market by a wide margin, with over 535 million users. For a great many customers it is not one messaging app among several — it is the default way they contact a business. A support channel that lives where the customer already is removes the step most likely to lose them.

The language picture matters just as much. More than 250 million Indians move between Hindi and English inside a single sentence, and published word-error rates for Hinglish speech recognition range from roughly 27% to 70% across monolingual-trained models. If your agent cannot follow a caller who switches language halfway through a request, the channel will not carry the volume you are hoping for.

Headline rates are the least useful number in this market. What determines your bill is which components sit inside the quoted rate. Published all-in rates from Indian vendors in 2026 land between roughly ₹3 and ₹10 per minute, and WhatsApp calls generally cost more per minute than the equivalent PSTN call.

The mechanical part is genuinely short — under an hour if your WhatsApp Business Account is already verified. The order is what trips people up.

The actual sequence

1. Have a verified WhatsApp Business Account. Nothing below works without it, and verification is the step with a human-review queue attached.
2. In the Meta developer console, enable the calling capability on the phone number you intend to use, then copy its Phone Number ID and your WABA ID.
3. Create a System User in Meta Business Suite and grant it the WhatsApp permissions. Use a System User token, not a personal one — personal tokens die when the person leaves.
4. Point the call webhook at your platform so inbound calls reach your agent, and verify the signature on every event.
5. Test the handoff path before the happy path. The transfer to a human is the part that fails in production.

Every vendor guide covers the setup. Almost none covers what happens afterwards, so here is the list worth planning for.

1. The handoff loses the thread

An AI agent that transfers a call without passing the conversation history makes the customer start over, which is worse than never having automated the call. Insist that whatever picks up second can see everything the caller already said — in the chat and on the call.

2. Mid-sentence language switches

A caller opens in English, gets frustrated, and finishes in Hindi. Agents configured for a single language per session mishear the second half entirely. Language detection has to run continuously, not once at the start of the call.

3. Concurrency ceilings arrive without warning

Concurrency is usually capped by plan, and the cap is invisible until a campaign or an outage sends everyone to you at once. Ask what happens to call number 51 when your ceiling is 50, and make sure the answer is a queue or a callback rather than a dropped call.

4. Identity does not carry across channels

The same person messaging on WhatsApp, emailing, and calling can easily become three unrelated records. If the channels do not resolve to one customer, your agent answers each contact with no memory of the other two — and your analytics triple-count the work.

5. Nobody disclosed the agent is a machine

Hiding that a voice is artificial risks being treated as a deceptive practice under Indian consumer law. This is a one-line fix at the top of the script and an expensive problem if skipped.

The channel is only worth opening if something useful answers. In practice that means an agent that can take actions, not just recite a knowledge base — checking an order, booking a slot, applying a credit — and that hands over cleanly when it cannot.

Adya today handles WhatsApp messaging, website chat, email and phone calls from a single shared inbox, in 10 Indian languages plus English, and follows a caller who switches language mid-sentence. When it escalates, the human takes over mid-conversation with the full history, which is the part that decides whether automation helps or annoys. It also books appointments with Google Calendar sync and runs outbound WhatsApp and AI-call campaigns.

Two frameworks apply, and they apply differently depending on how the call is carried.

Commercial calling rules

TRAI’s DLT registration and the TCCCPR framework govern commercial voice calls placed over Indian telecom networks, and they make no distinction between a human and an AI dialler. Outbound number series are prescribed — 1600 for service and transactional, 140x for promotional — with frequency caps and a 9am to 9pm window for promotional voice. Penalties are assessed per call, not per campaign.

WhatsApp calling travels over Meta’s infrastructure rather than the PSTN, so its treatment under these rules is not settled the same way. That is a question for your counsel before you run outbound at volume, not something to infer from a vendor blog.

Data protection

The DPDP Act, with rules notified in November 2025, requires explicit informed consent to record and process a call, and this applies regardless of which channel carried it. A customer who simply has not opted out of commercial calls under TRAI has not thereby consented under DPDP — they are separate permissions. You will need automated disclosure, a consent log, encrypted storage and role-based access to the recordings.