How do you build AI that makes the process of calling up customer support not…suck? Regal co-founders Rebecca Greene and Alex Levin know better than probably anyone: they’ve raised more than $80M to reimagine AI phone agents for contact centers.
In this episode of Barrchives, we discuss how Regal is building their voice agents, what makes deploying these systems so difficult, where speech-to-speech models need to improve, and where contact centers are going to be in 5-10 years.
“Let’s be really clear: anybody building AI assist for human agents, their business will go to 0 in the next 10 years. It will all be AI agents, plus managers who are human, and P&L owners who are human, and prompt engineers who are human. There will be humans, but they won’t be doing the tier 1 thing. I also believe it’s going to be more voice than text: humans talk faster than they type, they emote better over voice, so I think you’re going to see voice become the dominant channel again.”
"The first challenge is the speech-to-speech models - they're in beta for a reason. They go down quite frequently, the cost is too high. The promise of these models, which isn't there yet, is not only lower latency but being able to detect emotion. Agents are good at saying the right things for obvious emotions, but they can't mirror everything we're saying. They can't laugh if you're trying to make a joke."
"Testing is one of our biggest challenges, especially in voice. It's not just 'is the logic correct?' but 'does it sound good?' One area of the prompt may affect a different capability you thought was already working. We're building solutions around large-scale testing: how do I create different scenarios of customers and personas, role-play conversations through the AI agent, and verify it passes compliance, friendliness, and task completion criteria?"
“There’s still a lot of value in prompting, and people are still learning the best ways to prompt models to get the best outputs. We’ve focused on some core use cases in the contact center that we think agents can perform really well on – collections, qualifications, inbound triage – we’re not trying to boil the ocean. And then we’re getting really good on how the prompting needs to work to be both conversational, but then also complete the task that the agent is supposed to do. A simple example: how do you get a customer to engage? Should you announce that you’re an AI? And how do you do it? What language do you use?”
.svg)