AI Voice Agents Are Calling Your SIP Trunk

AI voice agents went from novelty to production infrastructure fast. Platforms like Retell AI, Bland, Vapi, and Voiceflow are handling real customer calls — appointment scheduling, order status, first-tier support, outbound campaigns. From the network's perspective, these are regular SIP calls. From a capacity planning and quality perspective, they are something entirely different.

If your organization is deploying AI voice agents, or if you are an MSP whose clients are, the SIP trunks and network infrastructure need to account for call patterns that human callers do not generate.

How AI voice agents connect

The architecture is straightforward. The AI platform runs a speech pipeline: a language model generates text responses, text-to-speech converts them to audio, and the audio is streamed into an RTP session over a SIP call. Inbound calls arrive on your SIP trunks and are routed to the AI platform. Outbound calls originate from the platform through your trunks to the PSTN.

From your SIP provider's perspective, it is a regular call. The INVITE looks normal. The RTP flows normally. The BYE terminates normally. There is no protocol-level difference between an AI agent call and a human call.

The differences are operational.

Call patterns that break Erlang assumptions

Traditional voice capacity planning uses Erlang formulas. Erlang B tells you how many trunk channels you need given a call volume and acceptable blocking probability. It works well for human callers because human calling patterns are predictable: they follow business hours, they have natural gaps between calls, they build gradually in the morning and taper off in the evening.

AI agents do not follow these patterns.

Instant scale-up. When an outbound campaign launches, an AI agent can attempt to place hundreds of calls simultaneously. There is no gradual ramp. One moment the trunks are idle, the next they are at capacity. Your SIP trunk provider may rate-limit or reject calls if the spike exceeds your contracted channel count or their fraud detection thresholds.

No breaks. A human agent handles a call, does post-call work, takes a break, chats with a colleague. An AI agent finishes a call and immediately starts the next one. Trunk utilization per agent is significantly higher than the equivalent human.

Uniform call duration. Human calls have variable duration — some are 30 seconds, some are 20 minutes. AI agent calls tend to cluster around a narrower duration range because the conversations follow scripted flows. This makes the peak concurrent call calculation different from what Erlang models expect.

Off-hours traffic. AI agents can call during hours when your trunk utilization was traditionally zero. If your capacity planning assumed daytime-only traffic, you may need to revisit the numbers.

How to size it

Forget Erlang for AI-heavy trunks. Instead, calculate directly:

1. Peak concurrent calls — what is the maximum number of simultaneous AI calls your system can generate? Include both inbound and outbound.
2. Headroom — add 20-30% above peak for human calls sharing the same trunks.
3. Provider limits — check your SIP trunk provider's concurrent channel limit and calls-per-second (CPS) limit. CPS limits are often lower than you expect — many providers limit to 1-5 new calls per second per account.

Use the bandwidth calculator to determine the throughput needed for your peak concurrent call count. Multiply concurrent calls by the per-call bandwidth for your codec (approximately 85 Kbps for G.711, 30-40 Kbps for Opus).

The codec question

Codec selection matters more for AI voice than for human voice, because synthesized speech has different acoustic characteristics than natural speech.

G.711 (PCMU/PCMA) — uncompressed, 64 Kbps. The safest choice for PSTN termination. No compression artifacts. The audio quality is limited only by the 8 kHz sampling rate, which is the same limitation human callers have on traditional calls. Works everywhere.

Opus — adaptive bitrate, wideband or narrowband. Excellent for AI voice because it can encode the full spectrum of synthesized speech and adapt to network conditions in real time. If both sides support Opus (common in WebRTC-based platforms), use it. The quality difference is noticeable — synthesized speech sounds more natural at higher bandwidth.

G.729 — compressed, 8 Kbps. Saves bandwidth but introduces compression artifacts. On human speech, these artifacts are minor. On synthesized speech, they can be more noticeable because TTS audio has different spectral characteristics than natural voice. The compression can make an AI agent sound more robotic than it should. Avoid G.729 for AI voice if bandwidth permits.

The codec negotiation happens in the SDP exchange during call setup. Make sure your SIP trunk provider and your AI platform agree on codec priority. If the provider forces G.729 and your AI platform is optimized for Opus, the quality will suffer.

Quality monitoring changes

Monitoring AI call quality requires the same metrics as human calls — jitter, packet loss, latency — but the interpretation is different.

Latency sensitivity is higher. In a human-to-human call, 150 ms of one-way latency is acceptable because humans naturally pause between conversational turns. AI agents respond faster — many platforms aim for under 500 ms response time including language model inference. Additional network latency stacks on top of that, and the caller perceives the total delay. If your network adds 200 ms of latency and the AI platform adds 400 ms of processing time, the caller waits 600 ms for a response, which feels sluggish.

Packet loss is less forgiving. Human callers unconsciously compensate for brief audio gaps — they fill in missed words from context. A caller listening to a synthesized voice does not extend the same benefit. Audio gaps in TTS output sound like glitches, not like a natural conversation artifact. Keep packet loss under 0.5% for AI voice, tighter than the 1% threshold for human calls.

Jitter buffer tuning. Some AI platforms manage their own jitter buffers. If the jitter buffer is too aggressive (large), it adds latency to the already-latency-sensitive AI response. If it is too small, jitter causes audio artifacts. Work with the platform vendor to tune the buffer for your network conditions.

Run our VoIP quality test on the network path between your infrastructure and your SIP trunk provider. The results show whether your network can support the tighter quality requirements that AI voice demands.

Regulatory and caller ID considerations

AI voice agents making outbound calls have the same regulatory obligations as human-dialed calls:

TCPA compliance. Automated calls to cell phones require prior express written consent. The FCC does not distinguish between an AI agent and a human agent for consent purposes. If anything, enforcement is stricter for automated systems.

STIR/SHAKEN attestation. Calls from AI agents go through the same STIR/SHAKEN framework. If your AI agent is generating high call volumes, your attestation level and caller reputation matter even more. A sudden spike in outbound calls from numbers that were previously low-volume can trigger spam flagging.

Disclosure. Several states and proposed federal rules require disclosure that the caller is an AI. Check your jurisdiction's requirements.

Use our carrier lookup to verify line types before outbound campaigns. Calling mobile numbers without consent carries higher regulatory risk than calling fixed lines.

Separating AI and human traffic

If your AI call volume is significant, consider separate SIP trunks for AI and human traffic. This gives you:

• Independent capacity planning. Scale AI trunks without affecting human callers' access to channels.
• Separate quality monitoring. Track AI call quality metrics independently. Problems with AI call patterns do not obscure issues with human calls and vice versa.
• Different provider configurations. You may want different codecs, different CPS limits, or different caller IDs for AI traffic.
• Blast radius containment. If the AI platform malfunctions and floods the trunks with calls, human callers are not affected.

This is the same principle as putting phones on a separate VLAN — isolate the traffic so problems in one domain do not spill over into another.

For SIP trunk fundamentals, see SIP Trunking: Connecting to the Outside World. For capacity planning, see Capacity Planning and Scaling.