AI Voice Agents and Telephony: SIP, WebRTC, and PSTN Explained
Deploying voice agents means connecting to the phone network. Here's a practical guide to SIP trunking, WebRTC, PSTN, and choosing the right telephony stack.
Your AI voice agent needs to connect to the world through phone networks or web browsers. The telephony layer is often the least understood and most frustrating part of deployment. Here's what you need to know about the three main connection types and when to use each.
PSTN: the traditional phone network
The Public Switched Telephone Network is what your customers use when they dial a phone number. PSTN connections are the most universal — every phone can reach them — but introduce the most latency (50–150ms round trip) and cost the most per minute. For inbound support lines and outbound calling campaigns, PSTN is typically required because that's how your customers make calls.
SIP trunking: the enterprise standard
Session Initiation Protocol (SIP) connects your voice agent to the phone network via internet-based trunks provided by carriers like Twilio, Vonage, or Telnyx. SIP is how most voice agent platforms handle telephony: they provision phone numbers on a SIP trunk, receive calls via the SIP protocol, and route audio to the AI pipeline. If you have existing SIP infrastructure (a PBX, contact center platform, or carrier relationship), you can often bring your own trunk and point it at the voice agent platform.
WebRTC: browser-based voice and video
Web Real-Time Communication (WebRTC) enables voice and video directly in the browser with no phone network involved. Latency is typically 20–50ms — significantly faster than PSTN. WebRTC is ideal for embedding voice agents in your website or app, where the customer clicks a button to talk rather than dialing a phone number. It also enables multimodal interactions (voice + video + screen share) that PSTN cannot support.
Choosing your deployment model
- Inbound phone support (customers call a number) → PSTN via SIP trunk
- Outbound calling (sales, reminders, collections) → PSTN via SIP trunk with caller ID management
- Website or in-app support (customer clicks to talk) → WebRTC for lowest latency and multimodal capability
- Hybrid (phone + web) → platform that supports both SIP and WebRTC deployments from the same agent configuration
Ready to build?
See how Mazed's multimodal AI agents work for your use case.