Phone AI Optimizes Voice Agents with Custom LLMs for 5% Gains
Phone AI's platform handles millions of calls monthly across verticals like insurance and home services, using custom LLMs and data analytics to boost outcomes by 5% via tweaks like changing one question, differentiating from basic voice AI.
Build Optimization Loops to Drive Voice AI Performance
Phone AI processes millions of calls monthly for call centers in insurance, home services, and hundreds of verticals, focusing on inbound leads from billboards and ads. Instead of just answering calls, the platform surfaces data to statistically prove improvements—e.g., changing one question lifted a customer's outcomes by 5%. Customers qualify leads, book appointments, or hand off to humans only when required (like licensed insurance sales). This data-driven loop lets businesses iterate prompts and flows, treating voice AI like e-commerce checkout optimization: capture all interaction data, analyze for conversion leaks, and refine. Result: AI handles 80% of calls indistinguishably from humans today, approaching 100% by year-end, reducing handoffs and prioritizing revenue-generating tasks over support.
Disclose AI for outbound calls due to emerging regulations, but inbound users prefer AI's context-aware, non-judgmental responses—no embarrassment asking 'dumb' finance questions or wasting human time on bookings. Trade-off: telephony nuances like garbled audio demand specialized handling beyond generic models.
Use Modular Custom LLMs to Cut Latency and Costs
Reject monolithic models like OpenAI's; break voice AI into specialist components—e.g., one for storing variables like names/emails, others for tasks—running on Groq's fast inference hardware. Benefits: slash latency (now 'good enough' like oxygen), lower costs, match quality, and isolate updates without retraining everything. Switch models dynamically by component, not per-question, enabling task-specific fine-tuning.
Bottlenecks shifted from latency to conversational quality, accuracy, interruption handling, endpointing, and edge-case transcription. Custom open-source LLMs, built from PhD experiments, provide battle-tested production reliability absent in off-the-shelf options. Future-proofing: deepen telephony expertise as generic models commoditize basics, but optimization platforms win on vertical-specific outcomes.
Bootstrap via SMB Feedback, Pivot to Enterprise
Start with $30-100/month SMBs for rapid feedback and iteration (4-5 months), then pivot when one call center outpaces all SMB revenue combined. Inspired by dad's small practice phone woes, evolved from basic receptionist to enterprise platform. Raised $16M Series A from Bessemer via LinkedIn post on ultra-endurance cycling lessons (300+ mile races teaching commitment).
Hiring in SF: sales, growth, engineering for low-ego, problem-focused team. Goal: scale to 50M+ calls/month. Founder advice: Expect endless daily battles—even top founders fight models/competitors; if you must found (not just 'want to'), roll dice relentlessly to create luck. Test fit at startups first.