SIP Trunking

SIP trunking is how phone calls travel over the internet instead of over copper. SIP is the signalling protocol that sets up, changes, and tears down a call; the "trunk" is the connection between your phone system and a provider that can reach the regular phone network. If your office phones are VoIP, you are already using it.

Why it matters when you are buying or building

For an AI voice agent, SIP is the layer that gets audio to and from the model. Most platforms give you two options:

• A number from the platform. Buy a DID directly from Telnyx, Twilio, or whoever runs the agent, and point calls at it. Simplest path, fewest moving parts, and what I use for most small business builds.
• SIP connection to your existing system. Keep your current phone provider and hand specific calls to the agent over a SIP trunk. More setup, but the client keeps their numbers, their extensions, and their existing routing.

The second option is what larger clients want, because ripping out a working phone system to add a receptionist bot is a bad trade. It is also where the integration work lives: credentials, IP allowlists, codec agreement, and which side is responsible for the media path.

What I have seen in production

Nearly every voice project I ship has a phone-plumbing phase that takes longer than the AI phase. Numbers on Telnyx and Twilio, SIP configuration, and forwarding out of whatever carrier system the client already runs. On one client I spent real time on a carrier's admin console before finding that group forwarding, not the built-in auto-receptionist, was the feature that would actually route calls to an external AI number. Nothing about that was an AI problem. See phone forwarding for AI agents.

Two things worth knowing going in:

Audio quality is a codec and network question. SIP will happily connect a call that sounds bad. Bad audio into speech-to-text produces bad transcription, and the agent then answers the wrong question. If an agent is misunderstanding callers, check the audio path before you blame the model.

Silence has many causes. Dead air on answer can be SIP media not flowing, or it can be a config bug in the agent. I have chased what looked like a carrier fault that turned out to be an assistant with empty dynamic variables, which produced total silence on answer. Knowing both layers is what makes that debuggable. See voice agent dead air debugging.

You do not need to become a telecom engineer. You do need to know which layer you are looking at when a call goes wrong.