YC Root AccessThe Voice AI Platform Powering a Billion Calls a Year
At a glance
WHAT IT’S REALLY ABOUT
Vapi’s founders built the infrastructure behind billion-call voice agents
- Vapi is a platform for deploying production voice agents and now serves roughly a billion calls per year for large customers like Amazon, Uber, and Intuit.
- The founders argue the biggest misconception is that voice AI is mostly a model problem, when in practice it’s an end-to-end production problem involving latency, cost, reliability, and business-process translation.
- Vapi emerged after years of pivots: a Zoom-join button and meeting note-taker reached meaningful revenue, but the team reset to pursue a more “inevitable” long-term opportunity in voice.
- A breakup-built AI therapist prototype clarified that real-time voice interaction would be transformative, and early customers like Hyperbound pulled the team toward building a scalable platform.
- Growth initially came from developer pull and intense support/iteration rather than marketing, then expanded into enterprise via contracts, compliance, and a more formal sales motion—without being displaced by AI labs.
IDEAS WORTH REMEMBERING
5 ideasVoice AI is hard mainly because production requirements (latency, guardrails, cost, reliability) are unforgiving.
Vapi’s founders emphasize that getting to production-grade voice agents involves mapping messy business processes into agent behavior, enforcing guardrails, meeting strict latency targets, and hitting acceptable costs—far beyond a demo that “kind of works.” Voice also has fewer constraints than chat, so reliability expectations are harsher.
A durable “10-year inevitability” thesis can outlast short-term model limitations and pivots.
They anchored their bet on the macro trend that models would become “cheaper, faster, better,” eventually reaching human-level interaction quality, even if the exact timing was unclear. That conviction let them persist through uncertainty and multiple pivots.
Walking away from a functioning business can be rational if it caps ambition and conviction.
Before Vapi, they built a Zoom-join button and then an on-device meeting note taker with strong retention and meaningful revenue, but concluded it wasn’t ambitious or “important” enough relative to where the world was heading. They shut down a working business, moved to SF, and reset.
A narrow, personal prototype can expose a repeatable platform problem when others share the same pain.
Jordan built a voice AI therapist during a breakup and became obsessed with making voice feel human (especially latency and natural turn-taking). That prototype revealed a broader platform need when early voice startups (notably Hyperbound) struggled with the same real-time constraints.
In developer infrastructure markets, fast iteration + great support can substitute for marketing early on.
Early growth came from extreme responsiveness: hundreds of support Slack channels, rapid feature shipping, and effectively no marketing or formal sales motion. Demand was pulled by developers building voice products who needed a reliable pipeline.
WORDS WORTH SAVING
5 quotesSo Vapi is a platform for deploying voice agents. Uh, today we have about a billion calls served, uh, annually, and we work with folks like Amazon and Uber and Intuit helping to, uh, automate their phone calls.
— Jordan Dearsley
I, I feel like when you're a founder who hasn't hit product market fit for so long, it becomes so elusive that it's not even a real thing. Like, you think of it like, like heaven. Like, you don't actually know.
— Jordan Dearsley
I remember one time I was talking to GPT-4, uh, and I had this long dialogue about my feelings and stuff like that, and it, and it was like, "You, you, you notice like you always talk about your feelings in the third person. Are you trying to distance yourself from them?" And I was like, "What the fuck?"
— Jordan Dearsley
We had about 300 Slack channels for support. And, and we would, within five minutes of someone giving us a request, we would ship the feature to production. And, and that's roughly... We, we, we did not do any marketing. We did not do any sales. We didn't do anything. We just built like, uh, a need that existed in the market.
— Jordan Dearsley
There's two generational problems in AI. One is AGI, and the other is just, like, deploying AGI in the enterprise. Both are equally difficult... and all of the labs have their A players on that first problem... and their B players on the second problem. You have your A players on the second problem. So that's the advantage.
— Jordan Dearsley
High quality AI-generated summary created from speaker-labeled transcript.
