AI Voice Agent Platforms

Testing and observability become the control layer for production voice agents

As voice agents move into production, reliability is the bottleneck. A16z podcast commentary notes that at 90% LLM reliability across 60 million calls a year, failures are unbearable in housing and healthcare, which is the segment EliseAI serves with voice. A distinct tooling layer is forming around this: Cekura (simulation, diagnosis, prompt rewriting), Noveum's NovaSynth (simulated callers at scale), Phonic (build, evaluate, observe), and ThunderPhone (built-in simulation testing and live monitoring). Quality assurance is turning into a standalone product category and a differentiator for the platforms themselves.

ThunderPhone offers phone agents at 2¢/min across 47 languages, and Dograh pitches a fully free open-source alternative to closed voice platforms with a visual flow builder and local deployment. Comcent open-sources SIP-based voice infrastructure, and Bolcho AI targets India with flexible LLM/STT/TTS swapping. Orchestration is becoming a commodity, pushing incumbents such as Vapi, Retell AI, Bland and Synthflow AI to compete on reliability, integrations and vertical depth rather than per-minute price.

Why it matters · Margins on horizontal voice-agent orchestration will compress, so investors should favor defensible data, distribution or vertical workflow ownership.

Voice agents are being packaged for specific workflows. Toma serves auto dealerships, Conduit serves hospitality in 140+ languages, Leaping AI serves trades and home services, Superunit handles employment verification calls, and HappyRobot AI targets supply chain. EliseAI, backed by an early bet on voice according to the a16z conversation, shows the vertical model working in housing and healthcare. Each wedge pairs voice with deep integrations (PMS, CRM) that horizontal platforms struggle to match.

Why it matters · Vertical integration and workflow data create stickier revenue than generic voice APIs, which makes these companies the more likely acquisition and scale-up targets.

The model layer is advancing quickly. Fish Audio's Drama 3 offers directable emotion and pacing across 83 languages, Eclatira and Ojin combine real-time voice with vision and expressive faces, and Chert targets video agents on FaceTime calls. ElevenLabs, at an $11B valuation, and Deepgram, AssemblyAI and Cartesia supply the speech stack underneath. The realism is already high enough that Sacca jokes a deepfake scam call is 'an ElevenLabs call for sure'.

Why it matters · Richer voice and video realism expands the addressable use cases but raises trust, fraud and regulatory exposure for deployers.

A new class of tools lets AI agents place and receive real calls. Dial provisions real numbers across 200+ countries, Kelam adds inbound and outbound calling for AI assistants, ClawCall navigates phone trees for cancellations, and Zinley gives agents their own phone number and email. Twilio and Dialpad remain the incumbent telephony reference points. Voice is becoming a general action channel for agents, not just a contact-center product.

Why it matters · Agent-native telephony APIs could displace legacy CPaaS usage patterns and open a new machine-to-business call volume.