The Developer's Guide to AI Outbound Calling That Converts

A production AI outbound calling system needs four layers: SIP trunking for telephony, an STT-LLM-TTS voice pipeline under 800ms latency, agent logic for qualification and scheduling, and CRM integration. Self-hosted Dograh has no per-minute platform fee, so you pay only for servers, AI models and phone minutes. Hosted platforms add their own fee, such as Vapi's $0.05 a minute.

Key Takeaways

• Self-hosting drops the platform fee. Servers, AI models and phone minutes still cost money.
• SIP trunking with open-source frameworks is the production-ready path.
• Dograh adds deployment plumbing that Pipecat alone does not provide.

Here we cover the system, the real cost and the voice stack, then take it to production in Dograh, the open-source AI voice agent platform we build.

What a Production Outbound System Looks Like

A production AI outbound calling system has four layers that need to work together: telephony, voice pipeline, agent logic, and integration.

Telephony means SIP trunking. Your voice agent connects to the phone network through a SIP trunk provider like Telnyx or Twilio. Real phone numbers calling real phones. SIP trunking is the most reliable foundation for production outbound because it gives you direct carrier access without proprietary middleware or per-call platform surcharges. Your trunk provider handles number provisioning, call routing, and carrier compliance. Your application handles everything else.

The voice pipeline is the STT-LLM-TTS chain. Speech-to-text converts what the callee says, a language model decides how to respond, and text-to-speech turns that response into audio. The full loop needs to complete in under 800ms to sound natural. Anything slower and callees start disengaging or talking over the agent. STT is usually the latency bottleneck, not the LLM.

Agent logic determines what happens during the call: lead qualification, objection handling, appointment scheduling, CRM updates. This is where business value lives, and it is also where most hosted platforms lock you in. They own the agent runtime, so modifying behavior means working within their constraints.

Integration ties everything to your existing systems. CRM writes after each call, calendar bookings for scheduled meetings, webhook triggers for downstream workflows, and analytics to track what is working. Skip the integration layer and your agents generate conversations that go nowhere. That failure is quiet, because the caller hangs up satisfied and the business finds out later. The same gap decides whether voice agents for restaurants and home services hold up once the call volume is real. Failed delivery rescheduling is the clearest case of this, where the agent has to pull live order state before it dials and write the chosen slot back during the call.

The Real Cost of Outbound Calling

Voice agent usage grew 9x in 2025, and four out of five companies plan to deploy voice AI for customer interactions this year. The market is moving fast. But most teams pick up hosted platforms without understanding what they actually cost at scale or what alternatives exist.

Every competitor guide talks about efficiency gains but none show the bill. Here is what AI outbound calling actually costs across the main approaches.

Hosted platforms charge per minute, and the headline rates are misleading. Vapi's base rate of $0.05/min does not include STT, LLM, TTS, or telephony. Your actual cost lands between $0.07 and $0.25/min depending on which models you pick. ElevenLabs Conversational AI charges $0.08-0.10/min on their Business and Pro tiers. Their LLM costs are currently absorbed but will eventually be passed through.

Self-hosted with Dograh means zero platform fees. You pay for your SIP trunk (around $0.005/min), your STT API (around $0.007/min), your TTS, and cloud infrastructure. AI calling systems reduce cost per interaction to $0.50-$5 compared to $5-$25 for traditional call centers, a 70-90% cost saving. But self-hosting drops it further. When calls cost $0.01-0.02/min instead of $0.10-0.25/min, you can run qualification calls, survey calls, and follow-up campaigns that would never make financial sense on a hosted platform. The economics of automation vs traditional systems tilt heavily toward self-hosted ownership.

Picking Your Voice Pipeline Stack

The voice pipeline is where call quality lives or dies. Each component has a latency budget, and the wrong choice in any slot drags down the whole system.

For STT, Deepgram Nova-3 is the default pick for most production voice agents. Streaming transcription at $0.0077/min with accuracy that handles phone-quality audio well. Their Growth tier drops to $0.0065/min. AssemblyAI works too, especially if you need built-in speaker diarization, but Deepgram's latency profile is better for real-time outbound where every millisecond counts.

For TTS, Cartesia Sonic 3 leads on latency at roughly 40ms time-to-first-byte and costs about one-fifth what ElevenLabs charges per character. When you are making hundreds or thousands of outbound calls per day, that pricing gap adds up fast. Deepgram Aura is worth considering if you want to consolidate STT and TTS on one platform, though Cartesia's voice quality edges it out for most use cases. For the LLM, any model with function calling works. Smaller, faster models tend to perform better for outbound because the conversations are structured and predictable. You do not need GPT-4-class reasoning for a lead qualification call. Most production outbound agents run on models in the 7B-70B range or on hosted APIs like GPT-4o-mini or Claude Haiku.

Pipecat orchestrates the full pipeline as an open-source Python framework that streams audio through STT-LLM-TTS with sub-500ms latency. It handles turn detection, interruption handling, and service integration across 50+ providers. AI-driven cold calling teams achieve 5-10% conversion rates compared to the 2.3% industry average, and personalized AI calls see a 36% higher meeting conversion rate. Pipeline quality directly drives these numbers. Slow responses, bad turn detection, or robotic TTS will tank your conversion rates regardless of how good your scripts are. For a deeper look at latency optimization, see our guide on speech latency in voice agents.

Open Source Alternative to Vapi / Retell

Self-hosted voice agent platform — no per-minute fees

dograh-hq/dograh

Star on GitHub

What production outbound needs that a demo skips

A demo dials a number and reads a script. A production system also needs a cost figure per call, an architecture decision you can defend, and a way to change providers without rebuilding the agent.

Sales teams using AI report a 50% increase in qualified leads, but that outcome does not require a $30k/year platform subscription. It requires a solid voice pipeline, good prompts, and reliable telephony, all achievable with open-source tools.

From Prototype to Production with Dograh

Pipecat gives you the voice pipeline. But going from a working demo to a production outbound system means building agent management, telephony integration, call scheduling, monitoring, and a way for non-developers to edit agent behavior. That is months of engineering work that has nothing to do with voice AI and everything to do with operational plumbing.

Dograh handles that production layer. It is open-source (BSD-2 license), fully self-hostable, and built on top of Pipecat. Dograh gives you the deployment and management infrastructure that Pipecat does not provide, the gap between a working demo and running 1,000 calls a day. The visual workflow builder lets your sales team design agent conversation flows with drag-and-drop. They can edit qualification questions, update objection scripts, or change scheduling logic without writing code. This matters when you are running outbound campaigns that need daily tweaking. Easy editing of live agents is one of the features teams ask for most, and hosted platforms make it unnecessarily hard by burying configuration behind API calls.

Built-in telephony integration supports Twilio, Vonage, Vobiz, and Cloudonix. Connect your SIP trunk, assign phone numbers, and launch outbound campaigns from the dashboard. Bring your own API keys for every component so you are never locked into a specific provider. AI voice agents now handle 50-60% of routine calls autonomously, lowering average cost per contact by 30-50%. With Dograh, that autonomous handling runs on your infrastructure, with your data, at your cost.

Getting Started and Staying Compliant

Install Dograh locally in under two minutes. The GitHub repo has a one-command setup. Connect a Telnyx or Twilio SIP trunk for outbound telephony, add your Deepgram and Cartesia API keys, and build your first outbound agent in the workflow builder. Test with a real call before scaling.

Before you scale up, handle compliance. Under the TCPA rules, an AI-voice call you place generally needs prior express consent. A sales call, one that encourages someone to buy or rent property, goods or services, generally needs prior express written consent under the FCC's calling rules; confirm your position with counsel. Scrub your lists against the National Do Not Call Registry. Log consent records with timestamps. Different states have additional rules on top of federal TCPA requirements, so check your target states before launching campaigns. The FCC has been tightening enforcement on AI-generated calls specifically, making proper consent documentation more important than ever. Self-hosting gives you an advantage here because your call data, recordings, and consent records stay on your infrastructure rather than a third party's servers. For regulated industries like healthcare and finance, this is often a hard requirement.

Gartner predicts 40% of enterprise apps will feature task-specific AI agents by 2026, up from less than 5% in 2025. Outbound calling is one of the clearest production use cases. The teams that own their stack now - telephony, pipeline, agent logic - will have a compounding cost and capability advantage as the market matures.