Opening the WhatsApp calling channel takes a few minutes. The real work is making sure the agent on the other end actually gets your customer's job done. This guide sets up an agent to answer incoming WhatsApp calls, step by step.
One shortcut up front: if you already have an agent running on a phone line, you barely need to redo anything. The knowledge base, conversation rules and handoff logic are shared. What follows is complete for anyone starting fresh, and a checklist for anyone who already has an agent.
1. Decide who answers on that number
The first decision is operational, not technical: who answers calls to this number?
Three common setups:
- The agent always answers. Simplest and most predictable. Calls never drop regardless of team load.
- Team during hours, agent after hours. A reasonable transition model for businesses whose team still handles phone calls.
- The agent answers and hands off when needed. This is the right one for most businesses. The agent picks up, understands the topic, resolves it if it can, and passes it to the team if it cannot.
To decide, calculate this: what is your current missed-call rate? Count after-hours calls and second callers who hit a busy tone. If that rate is above ten percent, having the agent always answer produces results faster than any transition model.
2. Keep the greeting short
On a voice call the first sentence matters far more than the first message in a chat. A long greeting bores the caller and delays them from stating their business.
The pattern that works has three parts: who you are, the recording notice, an open question.
"Good afternoon, this is Ada, [company]'s digital assistant. This call is being recorded for quality purposes. How can I help you?"
What to avoid:
- Implying the agent is human. Admitting it when asked is not enough — saying it up front is the correct move.
- Opening with a promotion. The caller called with a purpose; hear it first.
- Building menu logic. "Press one for orders, two for returns" is pointless with a voice agent; the agent understands the topic from natural speech.
3. Prepare the knowledge base for spoken language
Content that works in chat does not transfer as-is to a call. In chat the customer can read a long answer; on a call they cannot listen to one.
When preparing your knowledge base:
- Shorten answers. One paragraph per question. The agent can give detail on a second turn if needed.
- Write numbers to be spoken. Prices, durations and addresses must be intelligible when read aloud.
- Flag the top five questions separately. The majority of calls usually cluster around five topics; answering those five flawlessly is worth more than answering a hundred topics adequately.
- Teach it to say it does not know. The agent must not guess on something it is unsure of — it should hand off. That is the easiest antidote to hallucination in knowledge base gaps.
4. Write your handoff thresholds
The handoff decision should not be left to the agent's discretion; it should be written as rules. Thresholds that work well in practice:
- The customer asks for a representative a second time — let the agent try once more on the first ask, and hand off without argument on the second.
- The topic is cancellation, refund amounts or a complaint.
- The call exceeds the duration you set without making progress.
- There is clear anger or insistent repetition in the customer's voice.
- The agent has answered the same question twice and the customer is still not satisfied.
How the handoff happens must be defined too: transfer the call if the team is available, and if not, promise a callback and turn that into a task. An unanswered handoff is worse than no handoff at all.
5. Set up recording and consent
Recording on WhatsApp calls works as it does on phone calls — and the legal obligations apply identically.
To complete during setup:
- The recording notice included in the greeting
- A retention period defined, and expired recordings deleted
- Role-based limits on access to recordings
- Voice calls and recording covered in your privacy notice
We cover these in detail from the Meta policy and data protection angles in the 2 August post.
6. Test before going live
Once setup is done, call from your own number and walk through each scenario below. These are the situations that cause the most trouble in the field:
- The normal path. Ask your most common question; check the answer's accuracy and length.
- Interruption. Cut the agent off mid-sentence. Does it stop and listen, or push on to finish its sentence?
- Silence. Say nothing. Does it handle the gap well, or stack questions on top of each other?
- Out-of-scope question. Ask something not in the system. Does it invent an answer, or hand off?
- Handoff insistence. Ask for a representative twice. Does the threshold fire?
- Noisy environment. Call from a crowded place. Can it follow the conversation?
- Language switch. Change language mid-call.
- Recording check. After the call, verify the recording, transcript and summary landed on the customer record.
An agent that passes these eight can go live. Every item it fails comes back as lost customers in production.
What conversation pace to expect
The first question after setup is usually "is the customer waiting?" Measured on our phone line, the model composes its answer in 0.65 seconds (659 and 624 milliseconds). What the customer experiences adds audio travelling both ways and the detection that they have finished speaking. WhatsApp calls run through the same voice-to-voice architecture; we are not publishing a separate figure for that channel.
If waiting in your own setup is noticeably long, the cause is usually an over-long knowledge base or unnecessary verification steps — a setup setting, not the architecture.
If you would like to hear the agent yourself first, you can have it call you from the demo page.