SIP trunking for Voice AI Agent Platforms: How Audio Reaches the PSTN

SIP trunking is the connection that carries real phone calls between the public telephone network and a Voice AI agent platform. It replaces physical phone lines with an IP connection, so AI voice agents can answer, place, and scale inbound and outbound calls without fixed circuits.

Teams building AI voice agents often treat speech models and orchestration as the whole stack. The missing piece is how live caller audio enters and leaves that stack. SIP trunking is that path. It connects your agent platform to the public phone network over IP, so caller speech reaches the speech-to-text (STT), language model (LLM), and text-to-speech (TTS) pipeline behind the agent, and the agent’s reply travels back to the caller.

This article explains what SIP trunking is, how it delivers audio to an agent, the trunk capabilities that decide whether an agent works in production, how the major agent orchestrators connect, market context for 2026, and the criteria that matter when you choose a provider.

What Is SIP trunking?

SIP trunking carries phone calls over an IP connection instead of physical phone lines. The trunk connects your voice agent platform to the public switched telephone network (PSTN), so agents can answer inbound calls on real numbers and place outbound calls to real people.

Three pieces do the work:

• SIP protocol: Session Initiation Protocol sets up, modifies, and ends real-time sessions. The IETF’s RFC 3261 defines the signaling standard used across the industry.
• The trunk: the logical connection between your agent platform and the provider’s network. Providers model this differently. Some issue a single trunk that carries both directions, with separate settings for where calls arrive and where they terminate. Others, Plivo among them, create a separate trunk per direction. Either way you are configuring two paths: one for calls coming in to your agent, one for calls your agent places out.
• The provider: routes calls to and from the PSTN, provisions numbers, authenticates your platform, and applies fraud and quality controls.

Signaling and media travel separately. SIP negotiates the session, then the audio itself flows as RTP between the provider and your platform. That RTP stream is what your STT model transcribes and what your TTS model speaks into, which is why trunk configuration has a direct effect on transcription quality and conversation timing.

Capacity works differently than it did for legacy trunks. Instead of provisioning a fixed number of simultaneous call paths up front, usage-based providers meter per minute and scale concurrency elastically. The planning question for an agent deployment is therefore not how much capacity to buy in advance, but how quickly the trunk absorbs a burst when a campaign starts or a support queue spikes.

Demand for this connectivity keeps rising as organizations move voice workloads to software. Mordor Intelligence’s SIP trunking market report projects the global market to reach $85.07 billion in 2026 and $181.58 billion by 2031, at a 16.38% CAGR. For teams deploying AI voice agents, that growth reflects a simple architectural fact: agents still need a reliable path to real phone numbers and real callers.

How SIP trunking Delivers Audio to a Voice Agent

In a Voice AI architecture, responsibilities split cleanly. The AI layer handles speech recognition, dialogue, and synthesis. The trunk carries the audio to and from the PSTN and tells your platform that a call is starting.

A typical inbound flow looks like this:

1. A caller dials one of your business numbers on the PSTN.
2. The provider receives the call and sends a SIP INVITE to your platform’s SIP endpoint.
3. Your platform decides which agent should answer, based on the number that was dialed, the caller’s number, or values carried in the SIP headers.
4. Media flows as RTP into the agent pipeline. Speech-to-text converts audio to text, the language model generates the agent’s response, and text-to-speech returns audio.
5. The agent’s speech travels back through the same path to the caller.

Step three is the one teams underestimate. A single trunk usually carries many numbers, so something has to decide which agent answers which call. Platforms name this layer differently, but the function is the same: a routing rule that keys on the number dialed, the caller’s number, or a header value, and can start a different agent with different instructions for each. Without it you end up running a separate trunk per use case, which multiplies both cost and configuration drift.

The same INVITE that starts the call already carries useful context: the caller’s number, the number they dialed, and any custom X- headers your upstream systems attached. If the provider passes those through unmodified and the platform surfaces them before the first turn, the agent can greet a known customer by name instead of asking who is calling.

Speech recognition behaves differently on phone audio than it does in testing. Calls arriving over the PSTN are narrowband by design, and recognition accuracy drops on digits, spelled-out names, and similar-sounding consonants compared with the wideband audio from a laptop microphone. Codec negotiation happens automatically during the SIP handshake, so this is rarely something you configure. It is something to design around: build confirmation steps for account numbers and spellings, and fall back to keypad entry when a digit sequence has to be exact.

Plivo SIP trunking connects an agent platform to the PSTN in both directions. You point your platform at an outbound trunk’s termination domain to place calls, and point an inbound trunk’s origination URI at your platform’s SIP endpoint to receive them. Authentication is either an IP access control list or credentials-based digest authentication, and signaling and media encryption are a per-trunk setting. Teams that would rather not assemble an agent stack at all can instead build on Plivo’s AI Agents platform, which pairs the same telephony network with a managed STT, LLM, and TTS pipeline. Both paths use the same underlying carrier network, so the choice is about how much of the agent stack you want to own.

Key Concepts and Terminology

Precise vocabulary helps you size systems and read vendor proposals without guesswork.

Trunk direction. Inbound and outbound are configured independently, whether your provider exposes them as one trunk with two sets of settings or as two separate trunks. The practical consequence is the same either way: teams frequently set up the inbound path, assume outbound works, and discover otherwise on the first outbound campaign. Even an inbound-only agent needs the outbound path if it ever transfers a caller to a human.

Concurrency. Concurrency is the number of calls in progress at once, and it is the capacity figure that matters for an agent deployment. Usage-based trunks scale it elastically and bill per minute rather than capping it at a provisioned amount. Ask what the ceiling actually is, how a burst is handled, and whether concurrency carries a separate fee.

Authentication. Trunks authenticate either by IP access control list or by username and password using SIP digest authentication. This is a compatibility question, not just a security one: the provider must support the mode your agent platform supports. Credentials survive an egress IP change; allowlists do not.

Encryption. Signaling is protected with TLS and media with SRTP, and on most platforms encryption is off until you enable it at both ends. There is a meaningful difference between negotiating encryption when both sides support it and requiring it, because the permissive setting will quietly complete a call in the clear if the far end declines. Regulated workloads should require it.

Transfers. Handing a caller to a human depends on the trunk supporting SIP REFER, and support is not uniform. Some carriers require REFER and PSTN transfers to be enabled on the trunk before they work at all, and some restrict which destinations are permitted. Confirm what caller ID the human sees after the transfer and what happens when nobody answers.

Keypad input. Agents need DTMF in both directions: capturing digits a caller enters, and sending tones to navigate an IVR on an outbound call. Ask whether the provider passes DTMF out of band as RFC 4733 telephone events rather than as audio, since compressed codecs distort inband tones and cause missed or doubled digits.

Numbers and portability. Local and toll-free provisioning, plus number portability, matter for global agent rollouts. In the United States, the FCC’s number porting rules require simple ports to be processed in one business day, which helps teams move existing numbers onto a Voice AI platform without renumbering customers.

Compliance. Healthcare, payments, and privacy regimes change what your trunk and platform must support. HHS guidance on HIPAA and cloud computing makes clear that cloud providers that create, receive, maintain, or transmit ePHI are business associates and need a BAA. Payment workloads fall under the PCI Data Security Standard. Ask vendors for the certifications that match your vertical before you put agents on live traffic.

How Voice Agent Orchestrators Connect to a SIP Trunk

Most voice agent platforms do not run their own carrier network. They expose a SIP endpoint and expect you to bring a trunk. The shape of that connection is the same across platforms, even though the parameters differ.

On Plivo, which creates a separate trunk per direction, two trunks do the work:

• Inbound. You create an origination URI holding the orchestrator’s published SIP hostname, attach it to an inbound trunk, then point your phone number at that trunk. Only inbound trunks can be attached to numbers.
• Outbound. The outbound trunk gives you a termination SIP domain unique to your account, which you configure on the orchestrator side along with whatever credentials it will authenticate with.

All of them accept a bring-your-own trunk, and all support both IP allowlisting and digest credentials in some form. What differs is which method each one defaults to or recommends, what sits behind an enterprise plan, and how the inbound hostname is issued, so it is worth reading the specific guide rather than assuming:

• LiveKit issues each project its own inbound SIP URI on a project-specific subdomain, and supports region pinning, which its documentation describes as restricting network traffic to a specific geographical region. Region pinning is arranged through their support rather than self-serve, and enabling it turns off automatic failover to the nearest region.
• ElevenLabs recommends digest authentication and falls back to access control list authentication when no credentials are supplied. Allowlisting only works if the far end presents static IPs, and ElevenLabs offers static IP SIP servers on enterprise plans only, so digest credentials are the practical path for everyone else. Data residency covering the EU, India, and Singapore is likewise an enterprise feature. Its inbound hostname is shared across customers.
• Retell AI requires both an inbound and an outbound trunk before it will accept a number, even when you only intend to use that number in one direction. Agents bind per direction, and the number is imported once carrying the termination domain, credentials, and agent bindings together.
• Vapi uses a fixed regional hostname for inbound rather than a per-customer one, and its integration path uses an IP access control list with UDP transport. It offers two hosted regions, US and EU, with EU data residency gated behind enterprise sales, and does not support media anchoring, which rules it out for markets that require call media to stay in country.
• xAI exposes a single static SIP hostname shared by all customers, with the phone number carried in the URI itself, and requires you to pick exactly one authentication method: an IP allowlist of permitted source ranges, or digest credentials. Its SIP documentation does not address regions or data residency.

Platforms outside that list follow the same pattern, which is what the guide for other platforms covers.

Two details recur often enough to check first. Transport parameters must match on both ends, and a mismatch usually shows up as a call that connects with no audio rather than a call that fails outright.

The second is media anchoring, which decides where call audio is actually processed rather than merely where it is routed. Some countries require that media for domestic calls terminate on infrastructure inside their borders, India among them, and a platform that cannot anchor media in the required region cannot legally carry that traffic no matter how the trunk is configured. Support varies by platform, so if you plan to serve a market with residency rules, confirm both the provider and the agent platform can anchor media there before you buy numbers.

Transfers work at the SIP layer. The Plivo SIP trunking overview documents the REFER flow: the agent sends a REFER naming the transfer target, Plivo accepts it and dials that target, then bridges the caller once the target answers. If the transfer fails, the caller stays on the agent’s line rather than being dropped.

Market Context and Adoption Drivers in 2026

SIP trunking is no longer only a cost exercise. It is the telephony foundation under AI voice agents that must answer real phone calls at scale, and the market reflects where that capacity sits. Mordor Intelligence’s SIP trunking market analysis reports that North America commanded 62.15% of market size in 2025, while Asia-Pacific is forecast to grow at a 16.27% CAGR through 2031. That split matters if you run multilingual or multi-region agents: North America remains the largest installed base, and Asia-Pacific is the fastest expansion lane for new capacity.

Cost still drives adoption. The same Mordor Intelligence research finds savings between 25% and 65% over legacy PRI lines as the strongest economic pull for organizations that still run circuits. For teams building agents from scratch, the more relevant comparison is between providers rather than against a circuit they never had, and the deciding factors there tend to be concurrency behavior, transfer support, and geographic reach.

Other drivers commonly cited by SIP trunking analysts and providers include:

• Agent platform adoption: every voice agent deployment needs phone numbers and a route to the PSTN, which pulls trunking into projects that would otherwise be pure software.
• AI-optimized routing: platforms route calls to agents or humans based on intent, language, or queue depth, which increases the value of elastic capacity.
• SME digitization: smaller teams adopt metered connectivity instead of buying hardware they will outgrow.
• Global expansion: local numbers in new markets are often the gating item for launching an agent there at all.

Restraints still apply. Quality of service on the public internet can vary. SIP fraud, including toll fraud and unauthorized international calling, remains a real operational risk, and an agent that can dial internationally is an attractive target. The practical response is strict authentication, geographic dialing restrictions, monitoring, and a provider that can fail over routes when a path degrades.

Where Trunk Capabilities Shape Agent Use Cases

The trunk does not change dialogue design, but it decides whether an agent can use real numbers, transfer to a human, absorb a spike, and satisfy a recording rule. Each use case below leans on a different trunk capability.

Healthcare. Agents handle intake questionnaires, insurance eligibility checks, post-discharge follow-up, and triage routing. The trunk-level requirement is a clean escalation path: triage calls that exceed an agent’s scope have to reach a licensed human mid-conversation, which makes reliable SIP REFER behavior and predictable caller ID on the transferred leg more important than raw call volume.

Customer service. Automated qualification, order updates, and first-call resolution deflect repetitive volume while keeping a transfer path to a human when confidence drops. The trunk requirement is burst behavior. A product launch or an outage can multiply call volume within minutes, so what matters is how quickly concurrency scales and what the platform does with calls that arrive while the agent pool is saturated.

Education and HR-tech. No-show reminders, program lead qualification, and candidate follow-up depend on outbound calling to people who did not ask to be called right then. That makes answering machine detection the deciding capability, since most of these calls reach voicemail and an agent that talks over the greeting wastes the contact.

Retail and e-commerce. Order status, returns handling, and promotional outreach run at high outbound volume, where caller ID reputation determines whether anyone picks up. STIR/SHAKEN attestation, which signs outbound calls in the US and Canada, is what keeps agent calls from being flagged as spam before the conversation starts.

Financial services. Account inquiries, payment reminders, and collections require strict control of cardholder data. The PCI Security Standards Council’s guidance on protecting telephone-based payment card data applies across mail-order/telephone-order and VoIP channels, including how recordings handle sensitive authentication data. Out-of-band DTMF helps here, because keypad digits carried as telephone events rather than audio do not land in the recorded stream.

Benefits and Considerations When Choosing a Platform

Selecting SIP connectivity for AI voice agents is both a telephony decision and a platform decision. Evaluate them together.

Call routing to the right agent. Confirm you can route on the dialed number, the caller’s number, or a header value, and that each route can start a different agent with different instructions. This is what lets one trunk serve many numbers and many use cases.

Context at call start. Ask whether custom SIP headers pass through unmodified and reach the agent before the first turn. Header values that arrive asynchronously are worse than useless, because the agent will have already greeted the caller generically.

Transfer behavior. Verify that SIP REFER is enabled, what caller ID the receiving human sees, and what happens when the destination does not answer within the ring timeout.

Regional origination and media anchoring. Where a call is anchored geographically affects both latency and compliance. A provider that terminates inbound calls on a regional endpoint and originates outbound calls near the destination keeps media from crossing continents twice on every conversational turn. The same control usually satisfies data residency requirements. Plivo runs points of presence in seven locations across five continents for this reason.

Failure modes and diagnostics. Most launch problems are configuration mismatches, not network faults, and they surface as specific SIP responses. A 403 usually means credentials do not match between trunk and platform; a 503 usually means the signaling address is wrong. Make sure the provider exposes a call ID that also appears in your agent logs, so you can follow one failed call across carrier, platform, and application instead of guessing which layer dropped it.

Guardrails. Set a maximum call duration and a ring timeout on the trunk. An agent that loops, or a call connected to an open line with nobody on it, will otherwise stay up and bill minutes silently, because nothing errors.

Commercial model. Pricing structure matters more than the headline rate. AWS’s SIP trunking overview notes that some services use flexible pricing so you pay only for what you use, which is the model most agent workloads sit on: billed per minute, with concurrency scaling on demand rather than purchased in advance. Compare a provider’s current published rates against your own call volume and destination mix, since international rates vary far more between providers than domestic ones do.

Reliability. Look for redundant points of presence, automatic failover when an endpoint stops responding, and clear uptime practices. An agent that cannot receive calls creates the same customer pain as a downed trunk.

Compliance and security posture. Map certifications to your industry, review encryption options and access controls, and confirm whether a BAA is available for PHI workloads. Plivo’s security and compliance documentation is one example of the artifact set procurement teams should request from any shortlisted provider.

Operational fit checklist:

• Concurrency ceiling and burst behavior
• Inbound and outbound trunk configuration, including separate authentication
• SIP REFER support and post-transfer caller ID
• Out-of-band DTMF, in both directions
• Custom header pass-through and timing
• Number provisioning, porting timelines, and caller ID options
• Regional endpoints and data residency controls
• Recording and redaction controls
• Support path for fraud events and quality incidents

Plivo’s SIP trunking covers this list directly: inbound and outbound trunks authenticated by IP access control list or digest credentials, signaling and media encryption toggled per trunk, SIP REFER enabled by default for human handoff, a fallback URI per trunk when an endpoint stops responding, and geographic permissions that restrict which countries your agents can dial.

Common Misconceptions and Pitfalls

Several assumptions slow projects or send teams down the wrong architecture path.

“SIP trunking is the same as WebRTC or WebSocket streaming.” It is not. SIP trunking is how you connect to the PSTN with phone numbers and carrier routing. Browser or app streaming transports solve a different media path. Some platforms offer both, but they are not interchangeable.

“SIP inherently adds major latency versus other transports.” Poor network design and distant media anchoring add latency. The SIP handshake itself is not automatically the dominant delay in a well-built pipeline. Measure it rather than assuming: instrument the time from end of caller speech to first byte of agent audio, and correlate it with a call ID that appears in both carrier and agent logs, so you can see which leg actually costs the time.

“Every provider is equivalent.” Security depth, fraud tooling, global number coverage, compliance attestations, and transfer behavior vary widely. A low per-minute price with weak authentication is not a bargain. Channel price is the easiest thing to compare and the least informative.

“Concurrency is the only capacity question.” Load-test with realistic agent speech patterns, including barge-in, long monologues, and callers who go silent, and check what happens at the ceiling. A trunk that queues calls behaves very differently from one that rejects them, and callers experience the difference immediately.

“Compliance is the agent vendor’s problem alone.” If call audio or recordings include regulated data, the trunk, the recording store, and the model vendors may all sit in scope. Assign ownership early.

Avoid these pitfalls and SIP trunking becomes a predictable audio layer under your AI voice agents rather than a surprise bottleneck.

Frequently Asked Questions

How many concurrent calls does a voice agent deployment need?

Size for peak, not average. Estimate concurrent calls as expected calls per hour multiplied by average call duration in hours, then add headroom for bursts. On usage-based trunks that scale concurrency elastically and bill per minute, the practical question is how fast the trunk absorbs a spike rather than how much capacity you reserved in advance.

Can a voice agent transfer a call to a human?

Yes, using SIP REFER, but support is provider-dependent. Some carriers require REFER and PSTN transfers to be enabled on the trunk before it works, and some restrict which destinations are allowed. Verify caller ID behavior on the transferred leg and the fallback path when the destination does not answer.

How does a caller reach the right agent when one trunk carries many numbers?

The platform applies a routing rule to each incoming call, keyed on the number dialed, the caller’s number, or a SIP header value. Platforms use different names for this, but each rule can start a different agent with its own instructions, which is what lets one trunk serve many numbers and many use cases.

Which voice agent platforms can connect to a SIP trunk?

Any platform that exposes a SIP endpoint. Plivo publishes step-by-step integration guides for LiveKit, ElevenLabs, Retell AI, and Vapi, plus a general guide for other platforms. The setup shape is the same in each case, while authentication method and transport differ by platform.

What is the difference between an inbound and an outbound SIP trunk?

Inbound is the path calls take from your phone numbers to your agent platform’s SIP endpoint. Outbound is the path your platform dials through to reach the PSTN. How that is packaged varies by provider: some issue a single trunk that handles both directions with separate origination and termination settings, while others, including Plivo, create a separate trunk for each direction. The distinction matters less than the checklist, since you have to configure both paths regardless, and some agent platforms will not accept a number until both exist, even when you only plan to use one direction.

How do you secure a SIP trunk used by AI voice agents?

Authenticate the trunk with either an IP access control list or username and password digest credentials, then enable encryption at both ends, TLS for signaling and SRTP for media. Prefer requiring encryption over merely negotiating it, since a permissive setting will complete the call unencrypted if the far end declines. Restricting which countries your agents can dial limits exposure to toll fraud.

Conclusion

SIP trunking is the audio path that lets Voice AI agent platforms participate in real telephony. It carries caller media into your STT, LLM, and TTS pipeline and returns agent speech to the PSTN with the numbers, concurrency, and controls your operation needs.

When you evaluate platforms, look past per-minute price. The trunk decides whether calls reach the right agent, whether the agent knows who is calling before it speaks, whether it can hand a caller to a human, and whether audio crosses an ocean twice on every turn. Those capabilities separate a pilot that demos well from a deployment that survives production traffic.

Plivo SIP trunking is built for that path: two trunk directions with a choice of IP allowlist or digest authentication, TLS and SRTP as a per-trunk setting, SIP REFER for mid-call human handoff, automatic failover to a backup URI, and geographic dialing controls. If you already run on LiveKit, ElevenLabs, Retell, or Vapi, the integration guides walk through both trunks step by step. Start free and map your concurrency, transfer, and compliance requirements before the first live campaign.