Phony.ai

Most outbound AI calling stacks have six layers — and builders rarely see all of them until something breaks. This episode maps every layer, prices out four real stack options, and names the compliance gaps that turn a working demo into a legal liability.

Show Notes

Building an outbound AI calling system is surprisingly easy to start and surprisingly hard to understand. The gap between a working demo and a production-grade stack is usually invisible — hidden inside vendor pricing tiers, opaque orchestration layers, and a compliance surface that most tutorials skip entirely. This episode of Phony.ai tears the whole thing apart, layer by layer, using the deep-dive article on outbound AI calling stacks as its technical foundation.

The episode walks through all six infrastructure layers every outbound call touches — and the four realistic ways teams are assembling them in 2026:

  • The six layers defined: campaign and dialer logic, carrier and phone number, media transport, speech-to-text, text-to-speech, and the language model — plus the layer most demos skip: consent evidence and suppression-list scrubbing before the first dial.
  • Twilio + your own orchestrator: Solid SDK coverage and managed PSTN routing, but the agent layer is entirely your responsibility. All-in costs land between 10 and 18 cents per minute once STT, LLM, and TTS are added, and the published median turn latency of 491 ms doesn't include carrier variance or orchestration overhead.
  • Telnyx + your own orchestrator: Roughly half the carrier rate of Twilio's long-code pricing, with deeper SIP controls including SIPREC and codec selection — but a narrower SDK surface. At volume, the default 20-calls-per-second-per-IP limit can produce silent SIP 503 failures if pacing isn't handled deliberately.
  • All-in-one platforms (Retell, Vapi, Bland, Synthflow): Fast to launch — days, not weeks — with everything collapsed into a single billing line around 7–10 cents per minute. The tradeoff is layer opacity: STT and TTS choices are vendor-curated, LLM execution is off your infrastructure, and a quiet TTS tier upgrade can triple voice costs without changing the invoice format.
  • Fully custom SIP stack: Floor costs can drop below 3 cents per minute at steady volume using tools like Jambonz or FreeSWITCH alongside self-hosted models, but the engineering burden — SBC config, codec negotiation, DTMF handling, 3 a.m. packet captures — is entirely owned in-house.
  • Compliance as infrastructure: The FCC's updated TCPA revocation rules require opt-out keywords to be honored within 10 business days, STIR/SHAKEN A-level attestation affects whether calls survive carrier scrutiny, and any stack that can't produce an immutable consent receipt per call attempt is a liability, not a product. Teams handling sensitive outbound use cases — from number provisioning and call routing through to call recording consent — need this layer designed in from the start, not bolted on later.

The episode also covers the latency math that determines whether a call feels human: the full round-trip budget from end of speech to start of reply is under 800 milliseconds, and carrier media paths alone contribute 80–150 ms before the model loop is even counted. For more on how AI voice capabilities can expand responsibly beyond initial deployment, check out From Receptionist to AI COO: Expanding Your Voice Agent Safely. More from the Phony.ai blog and platform documentation is linked throughout the episode.

Phony.ai

What is Phony.ai?

AI phone and voice agents, explained through the constraints that actually decide whether one works: end-to-end latency and where it comes from, interruption and barge-in handling, telephony plumbing and call control, transfer design, and the disclosure and recording rules around automated calls.

Each episode takes one design decision and works through it concretely — provider-neutral, comparing approaches rather than selling one. Written for teams evaluating, buying or building voice AI who need to know what breaks before it breaks in production. Five or six minutes an episode.

Topics include end-to-end latency and where it comes from, barge-in and interruption handling, telephony and call control, transfer and escalation design, prompt and turn design, evaluation and call review, and disclosure, consent and recording rules.

Produced by Phony.ai, provider-neutral AI phone and voice agents. Full details, services and further reading at https://phony.ai