Phone capability for AI agents: low-latency, low-cost calling API

AI-agent developers and modern contact-center teams increasingly need reliable, low-latency access to the public switched telephone network (PSTN) to connect LLM-driven voice assistants to real-world callers, but current CPaaS offerings were designed for broad compatibility rather than optimized streaming to LLMs and often introduce unpredictable latency, jitter and per-minute costs. This problem affects startups building customer-facing voice agents, mid-market contact centers modernizing IVR/assistant stacks, and enterprises experimenting with voice-first UIs within a $30.0B global cloud voice and contact-center market (1,000,000 businesses × $30K ACV). You could build a developer-first calling API explicitly optimized for AI agents: WebRTC-native ingestion, packet-preserving media proxies, deterministic PSTN bridging, built-in low-latency streaming transcription and hooks for LLM context windows, with contractual latency targets (median RTT <200 ms) and usage-based pricing designed to be 30–50% cheaper than incumbent minute-based CPaaS rates. Expose simple SDKs and a sandboxed test-number fleet so developers can prototype phone-capable agents in hours rather than weeks. Now is an attractive time because LLM-driven voice assistants are moving from demos to production, CPaaS commoditization is shifting buyer focus to latency, cost and domain-specific tooling, and the addressable market of roughly $30B with high ACVs supports both self-serve growth and enterprise deals. You can stand out by delivering verticalized agent tooling (conversation-state APIs, tokenized billing for model+telephony usage, built-in compliance/recording) and verifiable SLAs that incumbents struggle to promise, but be realistic: carrier contracting, global compliance, and multi-million-dollar upfront infrastructure and operational costs are significant barriers that require 2–3 year runway and strong engineering and carrier partnerships to overcome.