LiveKit Head of Product on Betting on Voice AI and Leading Without Authority

What does product leadership look like when AI makes everyone a builder? In this podcast hosted by iDonate VP of Product and Engineering Nacho Andrade, LiveKit Head of Product Ben Cherry will be speaking on how LiveKit made a high-conviction pivot from video infrastructure to voice AI, on betting on voice AI, why making AI sound human is harder and more important than most people realize, and what he learned moving from software engineer to product leader. His take on the “PM as CEO of the product” framing: it’s about accountability for outcomes, not authority over people, and getting that wrong is one of the most common traps in early product careers.

Subscribe to the Product Talk podcast on Spotify and Apple Podcasts and catch every conversation with leading product executives. Episodes drop on Wednesdays.

Show Notes

1. LiveKit started as the leading open source and cloud platform for WebRTC, the underlying technology behind browser-based video conferencing. That technical foundation turned out to be exactly what voice AI needed — a real-time, low-latency protocol for talking to AI models — and it set up everything that followed.
2. LiveKit’s first AI customer was OpenAI, and their technology powered both the first version of ChatGPT Voice Mode and ChatGPT Advanced Voice Mode. That early partnership gave them an unusually clear view of where the space was heading before most of the market was paying attention.
3. The pivot was a genuinely bold bet. LiveKit had a growing, successful video conferencing and live streaming business when the founders decided to double down on voice AI. They changed their pricing model, refocused their product roadmap, and shifted nearly all their resources — even as most of their biggest customers were still on the video side as recently as last year.
4. The signal that made the bet feel right was the consumer and media reaction to ChatGPT Voice Mode. People have been dreaming about computers you can just talk to for decades — it’s in every sci-fi movie, every Steve Jobs keynote. When it finally worked and felt natural, the enthusiasm made it clear this was a wave worth riding.
5. Voice AI adoption has taken years to become mainstream, and the pivot required patience. The breakout customers in the space are only starting to emerge now. Moving with high conviction before the metrics fully support you is a different kind of product discipline than most teams practice.
6. The fastest-growing use case for voice AI is customer support. Businesses that have spent years moving call centers to South Asia are now moving them to data centers, and it is a straightforward cost and quality calculation. But the consumer applications — language learning, accessibility, personal AI — are where Ben sees the bigger long-term impact.
7. One of the most memorable things Ben has learned came from a developer who built a talking jack-o’-lantern for Halloween using a Raspberry Pi and LiveKit for about $85, used it for one month, and cancelled his account when Halloween ended. The barrier to building creative, bespoke software has dropped so dramatically that use cases with nothing to do with enterprise automation are now very real.
8. Making voice AI sound human requires solving a lot of small, specific problems. One example is back-channeling — the “uh-huh” and “yeah” responses people naturally make mid-conversation. Six months ago, any voice AI would interpret those as prompts and stop talking to formulate a response. LiveKit built a small dedicated model to recognize back-channeling and ignore it, which is the kind of invisible infrastructure that makes conversations feel natural.
9. What developers building on LiveKit most want right now is the expressivity and empathy of the newest speech-to-speech models combined with the control to customize voice, domain, tools, and behavior. Full duplex models like GPT Live are impressive but hard to steer. The demand is for that level of performance without the black box.
10. Ben is most energized by voice AI as an accessibility technology. There are enormous numbers of people who do not consider themselves fluent in computing, who have not learned to type well, who struggle with graphical interfaces. Voice removes those barriers entirely and opens up everything the information economy has to offer to people who have been largely shut out of it.
11. LiveKit is also investing in robotics, and Ben sees a direct connection. Their real-time media transport network is already used by robotics companies for teleoperation and training observation, and voice is the natural interface for any robot operating in the real world. The same technology stack that powers voice AI is what robots will need to communicate with humans.
12. Ben draws an analogy to coding AI: for a long time the pieces were all there but the model wasn’t good enough, and then suddenly it flipped. He believes robotics is approaching the same inflection point, and when it flips, it will feel sudden even though years of work will have made it possible.
13. Ben’s path from software engineer to product leader was not a planned transition. He always gravitated toward user problems rather than pure technical challenges, and after years as an engineer he realized the questions he cared most about — which problems to solve and how they create value for users — were product questions, not engineering ones.
14. He spent a year or two at his previous company deliberately not writing code, trying to understand what a PM actually does between the conversations with engineers. His conclusion: the job is fundamentally about figuring out which problems to solve and then figuring out how to solve them, and that is harder than it sounds.
15. The most misleading piece of career advice for aspiring product managers is that the PM is the “CEO of the product.” The accountability part is true — the PM is accountable for whether the product succeeds. But the authority framing is wrong. Engineers don’t report to you, designers don’t report to you, and assuming you are the boss leads early PMs badly astray.
16. Influence in product comes from soft power, alignment, and helping people stay focused on the right problems. The moment you start telling people exactly how to do something, you lose ownership from the team. The job is to assign outcomes, not tasks, and then trust people to figure out how to achieve them their way.
17. One of the hardest things Ben has learned is to resist the urge to solve the problem himself even when he knows how. Telling someone “just do this” almost never works the way you want — not because they are wrong, but because they need to own the solution in their own way. Empowering that ownership is what produces the best results.
18. Ben does not believe in detailed roadmaps. You need a direction and a clear sense of which problems you are trying to solve and for whom, but a list of features mapped to specific quarters is usually wrong before the quarter even starts. Knowing what matters now and what comes next is more valuable than a plan that will need to be thrown out.
19. The product leader’s job is to stay one step ahead of the builders. If you get too deep into the details of how something is being built, you lose sight of what is coming next, and the team loses direction. Someone always needs to be looking around the corner, and that is the irreplaceable part of the role regardless of how capable AI becomes.
20. AI is already making Ben faster at the research and synthesis work that feeds good product decisions — competitor analysis, customer interview synthesis, early-stage exploration. But he is deliberate about writing the final artifact himself. If you let AI write the conclusion, it wastes everyone’s time because it never quite captures the actual thinking that needs to happen.