For years WhatsApp was a messaging channel for businesses. Voice calling existed on personal accounts but not on the business side; a customer calling a business number either got an error or simply no answer. The Calling API opened that side up: a business number can now be called from inside WhatsApp, and the business can initiate calls from inside WhatsApp too.
This post covers what the channel is, what it solves, and where its limits are — because the most misunderstood question when deciding on a rollout is whether this replaces your phone system.
What happens technically
WhatsApp calls are voice conversations carried over the internet. No GSM network, no line, no minute quota, no carrier bill; the audio travels over a data connection. On the business side, what answers the call is not a handset but a service that receives the call programmatically.
In practice the flow works like this:
- The customer starts a call from the business profile in WhatsApp.
- The call arrives as a notification to the service attached to the business number.
- The service accepts the call and a two-way audio stream is established.
- On one end of that stream is the customer, on the other your voice agent.
- When the call ends, the recording, transcript and outcome are written to your system.
Nothing changes on the agent side. However the agent behaves on a phone line, it behaves the same here; only the pipe carrying the audio is different.
What it solves
It zeroes out international calling costs. A customer abroad pays no carrier fee to call you. For exporters, businesses with overseas customers, or anyone in tourism, this means calls that never used to come in start coming in.
It lets the customer call from where they already are. They do not leave the app to find you in their contacts. They type on the same screen and they call from the same screen.
It connects the call to the context of the conversation. Who is calling, what they wrote before, which order they have — all known in the first second of the call. On a traditional phone call you face a number; here you have a customer record.
It reduces number verification friction. Because a WhatsApp account is already tied to a number, the caller's number matches the number in your record.
What it is not
Being precise here matters, because a wrong expectation produces post-rollout disappointment.
It is not a phone system. Extensions, call queues, transfers between team members — those classic PBX capabilities are not this channel's job. The party that builds those flows is your platform, not WhatsApp.
It does not replace your landline number. Some of your customers will always call you the traditional way. The Calling API does not remove that; it adds a channel beside it. The right setup connects both to the same agent rather than switching one off.
It is not a free-for-all outbound marketing channel. Meta does not allow businesses to call users at random; outbound calls happen within user permission and specific rules. Treating this as a channel for burning through a cold call list would be a mistake — and it would not produce business results anyway.
It does not offer an unlimited, unchanging rule set. Meta's business messaging and calling policies are updated regularly. When rolling out, check Meta's current documentation for permission requirements, call initiation conditions and country availability. The framework in this post is structural; policy detail can change.
Audio quality and latency
Audio carried over the internet is sensitive to connection quality. On weak mobile data a WhatsApp call can degrade; the GSM network is more predictable in that respect. In exchange, audio bandwidth is better than GSM in most cases, so when the connection is good the voice is clearer.
On latency, the deciding factor is not the network but the architecture of the model on the other end. In chained systems that convert speech to text, text to model, and model back to speech, every step piles on delay. In a voice-to-voice model those steps do not exist.
Measured on our own phone line, the model composes its answer in 0.65 seconds (659 and 624 milliseconds, two independent measurements). The same architecture runs WhatsApp calls, but we have not published a separate measurement for that channel. In chained systems every conversion step adds its own share, so total waiting grows; that is where the difference we observe in the field comes from.
Questions to answer before rollout
Opening the channel is technically easy. The hard part is deciding what happens once it is open:
- Will the agent answer incoming calls, or the team during certain hours?
- On which topics, if pressed, should the agent hand off to a human?
- Will calls be recorded, and in what wording is the recording notice given?
- In which scenarios will outbound calls be made, and how is permission obtained?
- Which CRM field does the call outcome get written to?
The answers to those five questions determine whether the channel works. In the next post we walk through setting up an agent to answer incoming calls, step by step.
You can see how channels come together in a single inbox on the omnichannel communication page.