Learn how to plan, connect and test an AI phone agent on a Twilio number, from inbound call flows to outbound batch campaigns, without guesswork.
Customers who message you on your website, WhatsApp or Messenger still pick up the phone when something is urgent or complicated. If nobody answers, they hang up and often do not call back. This guide explains how an AI phone agent works on top of Twilio, how to design a call that sounds natural, and how to test it before real callers reach it.
Why the phone still deserves an AI agent
Chat is easy to automate because people type short, clear questions and can scroll back through the answer. Phone calls are different. Callers ramble, interrupt, spell out names and expect a reply within a second or two.
That is also why the phone is where an AI agent can help most. Many calls are repetitive: opening hours, booking status, delivery questions, pricing, directions. A human who is already busy with a customer cannot pick up all of them.
An AI phone agent is a good fit when:
- Calls arrive outside staffed hours, or in bursts you cannot staff for.
- The same dozen questions account for most of your calls.
- You already have a knowledge base that your chat agent uses.
- You need a first response, not necessarily a final resolution.
It is a poor fit when most calls are emotionally charged or legally sensitive. In those cases, use the agent to collect details and route to a person, not to resolve the case.
What Twilio does, and what it does not
Twilio supplies the telephone side: a number that people can dial and the connection between the phone network and your software. The AI agent supplies the conversation: understanding the caller, finding the answer and speaking it back.
In practice, the pieces look like this:
- A caller dials your Twilio number.
- Twilio sends a request to a webhook address that belongs to your AI platform.
- The platform answers with instructions, and the agent starts the conversation.
- Speech from the caller is turned into text, the agent works out a reply, and the reply is spoken back.
Two details matter more than they seem. First, the webhook must be reachable and correctly registered, or the call fails silently from the caller's point of view. Second, requests should be verified as genuinely coming from Twilio, so that nobody can trigger your agent by sending forged requests. Good platforms check the signature on each incoming request.
Designing a conversation that works by voice
Text answers do not translate well to speech. A chat reply with a bullet list and a link is useless when read aloud. Plan the voice persona separately from the chat persona.
Keep answers short
Aim for one or two sentences, then a question or a pause. Long answers lose the caller, and they cannot skim ahead.
Confirm what matters
Names, email addresses, dates and phone numbers are where speech recognition makes mistakes. Have the agent repeat important details back: "I have Friday at three o'clock. Is that right?"
Plan the greeting
Say who is speaking, that it is an automated assistant, and what it can help with. Callers are more patient when they know what they are talking to, and a clear opening reduces the number of people who simply say "agent" and wait.
Decide the exits
Every call flow needs a way out. Decide in advance what happens when the agent does not know, when the caller is frustrated, or when the question is outside scope. Typical exits are taking a message, collecting contact details for a callback, or ending politely with the right opening hours.
- 01Caller dials number
- 02Agent greets and asks purpose
- 03Answers from knowledge base
- 04Confirms or takes details
Inbound and outbound are different jobs
The same Twilio account can support both directions, but the design thinking differs.
Inbound: answer first, then collect
Inbound calls reward breadth. Feed the agent the same material your support team uses: FAQs, policy pages, product documents. Then limit what it may promise. An agent that says "I will check with the team and someone will call you back" is far better than one that guesses a refund policy.
Outbound: narrow and respectful
Outbound batch calls suit appointment reminders, delivery confirmations, simple surveys and follow-ups with people who have already dealt with you. Keep the script short, identify yourself immediately, and give the person an easy way to end the call.
Consent and calling rules differ by country and by purpose, and they change. Check the rules that apply where you and your recipients are located, and get advice from a qualified professional if you are unsure. The platform cannot make that decision for you.
Test before anyone real calls
Most phone failures are found in the first ten test calls. Do not skip them.
A pre-launch checklist
- Call the number from a mobile phone, not just from a browser-based test.
- Call from a noisy place, such as a street or a café, to see how recognition holds up.
- Try an accent or speaking pace different from your own.
- Interrupt the agent mid-sentence.
- Go silent for ten seconds and see what happens.
- Ask something that is not in the knowledge base and check that it admits it.
- Ask for a human and confirm the agent responds sensibly.
- Read a phone number aloud and check that it is repeated back correctly.
For outbound campaigns, add a further rule: always send yourself a test call before the real batch, and know how to stop the run if something sounds wrong.
Language and voice choices
If your customers speak several languages, voice coverage is worth checking early. Speech recognition and speech synthesis do not always support the same languages equally well, and some languages may use a fallback voice. Test each language you plan to support with a native speaker rather than assuming parity with text chat.
Voice quality is also a brand decision. Some businesses want a neutral, calm voice; others want something closer to a named character. Pick one, keep it consistent between greeting and answers, and avoid unusual voices for serious topics such as billing or health.
Measuring and budgeting phone volume
Phone minutes cost more than chat replies, so estimate before you scale. The numbers below are purely hypothetical, to show the method.
In this made-up example, 60 percent of calls end with the agent resolving the question. That is the figure to watch over time. Review a sample of the "abandoned early" calls, because they often reveal a greeting that is too long or a question the agent answers badly.
Track these over the first month:
- How many calls reach the agent at all.
- How many end with a clear answer or captured details.
- How many callers ask for a person.
- Average call length, since long calls cost more and often signal confusion.
Then adjust the knowledge base and persona, not just the voice.
Doing this in AI SNS CHAT
AI SNS CHAT connects a Twilio number to an agent and handles both inbound answering and outbound batch calling. Everything lives under the Voice section of the sidebar, which includes Overview, Voice agent, Voice Playground, Inbound calls, Call history, Outbound calls and Voice analytics.
Inbound calls
- Create or choose the agent that should answer the phone, and make sure its knowledge sources are in place under Data sources.
- Open Voice → Voice agent to set the voice and speaking behaviour. Text-to-speech options include Grok/xAI (the default), OpenAI and ElevenLabs, and ElevenLabs adds a voice library, voice design and cloning.
- Try the persona in Voice Playground before connecting a number.
- Open Voice → Inbound calls and connect your Twilio number to the agent. Each agent gets its own number connection.
- Run the webhook self-test. If something is misconfigured, auto-repair can fix the webhook for you.
- Call the number from your own phone and review the result in Call history.
Webhook signatures from Twilio are checked, so forged requests are not accepted as real calls.
Outbound calls
Under Voice → Outbound calls, you can run batch campaigns from a Twilio number. A test call is required first, and a running campaign can be stopped midway. Keep the script narrow and make sure you have a lawful basis to call each recipient.
What to know before you rely on it
- Voice usage is billed in credits per voice minute, and the rate is set by the platform. See pricing for the current plans.
- The voice stack covers 22 languages, and some use an English phone voice as a fallback. Test each language you intend to serve.
- A SIP trunk section exists in settings, but a Twilio number is the supported route for calls today.
- WhatsApp voice and managed phone numbers are listed as coming soon.
- Voice analytics sit alongside chat analytics, though advanced analytics need the Pro plan.
For escalation, the handoff rota under Helpdesk → Scheduling assigns handoffs to whoever is on shift and can alert Slack, Teams, LINE WORKS, WhatsApp, email or a webhook. Helpdesk needs Standard or higher. See features and security for more detail.
Frequently asked questions
Do I need my own Twilio account?
Yes. You connect a Twilio number to your agent, and outbound batch campaigns also need a Twilio number. Twilio's own call charges are separate from your AI SNS CHAT credits.
Can the same agent handle chat and phone?
The agent's knowledge sources can serve both, but you should adjust the persona for voice. Spoken answers need to be shorter and free of lists and links, so test them in Voice Playground first.
What happens if the webhook breaks?
The Inbound calls page includes a webhook self-test and auto-repair. Run it after changing numbers or settings, and then place a real test call to confirm.
Can I use my own SIP trunk instead of Twilio?
The settings can be saved, but no call path that uses them is available at the moment. Use a Twilio number for now.
Is phone calling suitable for marketing campaigns?
Outbound calls are best for narrow, expected messages such as reminders and follow-ups with existing customers. Calling rules and consent requirements vary by country, so check what applies to you before you send a batch.


