When Jeff Pulver invited me to attend Voice and Conversations on the Net (VON) 2026 on behalf of WebRTC.ventures, I accepted immediately. Three days with people building the next generation of voice communications was an easy yes.
While at the conference earlier this week, the enthusiasm for voice AI was clear, and it extended well beyond the companies building the agents. Telecom providers, infrastructure engineers, startup founders, and enterprise teams were all part of the conversation.
Across all three days, I had great conversations both during the sessions and between them. That mix was a big part of the value: people approaching the same opportunity from different directions, with plenty to learn from each other.
Voice AI and telecom have a lot to learn from each other
The October 14 program captured that mix. Jeff Pulver opened with “The RenAIssance of Voice.” Randy Layman, CTO at AVOXI, had a more pointed question in “Can the Bot hear me?” It was a useful pairing: enthusiasm about what voice AI makes possible, alongside the fundamentals that make a conversation work.
Jonathan Eisenzopf of Vapi presented “VoIP 3.0: Conversations Not Call Legs,” while Kishor Patil of Best Buy brought an enterprise perspective where he demonstrated what omnichannel communication can look like with an agent assisting in a flow to reschedule a flight. The program also included Mike Mills of Gamma and Dag Peak of Alianza, connecting the discussion to the service-provider side of the industry.
For me, that is what makes this an interesting moment for telecom. An AI agent still needs a usable audio path, reliable call handling, and a way to recover when something goes wrong. Meanwhile, a connected call is only the beginning of what an enterprise needs: the agent has to understand the request and do the right thing with it.
At WebRTC.ventures, we work where those responsibilities meet. Seeing the telecom and AI communities compare notes in the same room was encouraging. Neither side has to solve the entire problem alone.
The conversation does not end when the call does
October 13 talks included the vCon Summit and VoiceAI LIVE!, bringing conversation data and working voice AI demonstrations into the same event. Thomas McCarthy-Howe of vConic was part of the program exploring how we capture and use those conversations.
The question is worth asking: after an agent finishes a call, what do we actually have to work with? A recording? A transcript? Enough context to understand what the caller requested, what the agent did, and whether the result was correct?
vCons provide a structured, portable container for conversation data. They do not, by themselves, prove that an agent behaved correctly, but they can help organize the evidence needed to investigate and evaluate it. We recently explored this with Jeff on WebRTC Live #118: Building the Data Layer for Voice AI with vCon. It was a useful lead-in to the conversations in Atlanta.
How do you know your voice AI actually works?
I also participated in the panel “How Do You Know Your Voice AI Actually Works?” That question connects the excitement around voice AI with the responsibility of putting it in front of real users.
My perspective is that a convincing demo is a starting point, not a release criterion. A caller can interrupt, correct an address, pause to think, or change a request while the agent is already doing something. An evaluation needs to account for that behavior, not just check whether the final transcript reads well.
Speed matters, but faster is not always better. An agent that starts answering before the caller finishes can look responsive while getting the interaction wrong. The useful question is whether the system behaves appropriately for the task and the person using it.
My talk: evaluating real-time voice agents beyond AI models
My presentation, “Evaluating Real-Time Voice Agents Beyond AI Models,” took that systems perspective further. Better models, prompts, and retrieval can improve an agent, but they do not replace the application’s responsibility for permissions, state, and recovery.
I organized the evaluation around four areas: conversation, comprehension, conduct, and completion. In plain language: does the interaction work, did the agent understand, did it follow the rules, and did the intended action actually happen?
I also introduced an open-source evaluation workbench and a practical workflow: run scenarios, evaluate the evidence, and compare results when the system changes. vCon helps make that evidence portable. The goal is to assess the whole system, rather than assume a better model score means a better production experience.
You can view my presentation here. I will cover the framework, examples, and evaluation workflow in more detail in a separate blog post.
From voice bots to wine bots
Automation was not limited to the conversations. I also came across a self-service wine bot: a lighter reminder that we are seeing more everyday interactions become self-service experiences.
After spending the day talking about automated voice conversations, seeing an automated pour felt fitting. A different kind of bot, and a good excuse for a photo.
Thank you, Jeff and the VON team
The talks gave us plenty to think about, but the conversations across all three days were just as valuable. I left with new connections, ideas to explore, and a renewed appreciation for bringing people with different kinds of voice experience together.
Thank you to Jeff Pulver for inviting us, and to the entire organizing team for coordinating the event and making those connections possible. It was a pleasure to be part of it. We are already looking forward to the next one next year!
At WebRTC.ventures, we help teams build, integrate, and test real-time voice AI applications, connecting the AI experience to the communications infrastructure behind it. Let’s talk about what you are building.
