Meet James Zammit, Co-Founder and CEO at Roark, who joined the BotCast #12 to answer our usual questions on building with Voice AI.
Who are you?
I'm James, one of the co-founders of Roark. Roark is a simulation and eval platform for voice agents. With Roark, teams can simulate thousands of conversations with different personas and conditions. We measure things like whether the agent understood the caller, interrupted them, followed the right flow, or actually achieved the goal of the call. Then we monitor those same things in production.
How did you end up using voice AI?
It's quite funny. My co-founder Daniel and I were really obsessed with home automation, so we started looking into building voice agents for that. We ended up building them for other people as well: dental clinics, car dealerships and a few others. We soon figured out they were really hard to test, and out of that we built this little simulation tool and got into the space. We've just been obsessed with speaking to a machine and getting it to turn on our lights or help us around the house.
A moment of joy and pain using voice tech.
Joy is simple: for us, it was when we won the Blabla Award at the London voice event you hosted last year. No, joking aside, I think the moment of joy and the moment of pain are the same. It's when the voice agent does something completely unexpected.
Sometimes it's in a good way, and it answers something you just didn't know it could figure out. Agents are getting so much better now. Whether it's reminding your grandparents to take their pills, helping them book a taxi, or all sorts of other things, it's just fantastic.
But at the same time, the pain is the unexpected. Sometimes they fail in ways you just don't expect. They're fully non-deterministic, so you're debugging several systems. Was the problem the speech-to-text? Was it the LLM? Was it the agent misunderstanding the question? That can be pretty painful.
What lesson would you share with other builders?
This one's pretty close to us: don't test only the happy path. When you're building, you ask, "Can this agent book an appointment?" That's the happy path. It sounds obvious, but with voice agents it's incredibly easy to keep testing them as yourself. You know how the agent works, you speak clearly, you phrase things in ways you know it will understand.
But real callers interrupt. They change their mind halfway through a sentence. They have different accents. They're in a car, they're annoyed, sometimes they barely even answer the question. One thing we've learned is that you need variation and edge cases very early.
Where do you see the industry in 12 months and in 5 years?
Over the next 12 months, I think the biggest change, and we're already seeing it, will be that the conversation shifts from "Can voice AI do this?" to "Can voice AI do this reliably?"
Over the next five years, and we said this even a year ago, we believe every publicly facing phone number will be a voice agent. Every company that has a publicly facing phone number today will just have a voice agent.
The interesting part to me is that voice probably stops being thought of as a separate category called voice AI, and just becomes another interface for software. It's no longer "this is a voice agent platform for X, Y or Z". It's a feature of software. You have your CRM, and obviously you have a voice agent on it, just like you have a chat widget or a way to book appointments.
Once that happens, the bar becomes really high. These systems are handling millions of conversations and increasingly doing important tasks, so companies will need testing, monitoring and infrastructure around these agents. That's what we do.
ROARK
https://www.voiceaispace.com/tool/roark-ai
BotCast powered by https://bespoken.ai 🫶
