Kubrio is one of the projects I'm working on: a learning app where kids build real things, so the people writing to us are parents. We're early, with 10 to 50 messages a week, and we answer most of them the same day.
Customers write tickets, and Claire answers
I moved support for customers to tickets only. An agent we call Claire manages them. Claire runs as a Grok Bot: every new ticket starts a run, with the ticket as its input. When a question is easy, Claire answers it on her own from our knowledge base. I see every answer and can step in almost in real time.
Right now Claire handles up to 10% of requests by herself. That's on purpose: I want to stay in the loop and learn what kinds of problems parents actually have. As the knowledge base grows, I plan to hand her more, until she covers every ticket.
Refunds are straightforward, so Claire handles those too. I take over when there's a real issue: a bug that hurts the experience, or anything Claire can't explain.
Everyone else calls, and a voice AI screens
For people who aren't customers yet, I moved everything to the phone. A voice AI answers like a PA: it asks who's calling and why they want to connect. If it's important, I call back.
The voice runs on ElevenLabs with its v4 voice model, and the phone line runs on Twilio. ElevenLabs lets you pass the agent context in real time during the call, so it can have a better conversation than a fixed script.
The plumbing
- Claire and the voice agent share one knowledge base about Kubrio.
- Email goes through AgentMail. When someone replies to our newsletter, the reply goes through Resend instead.
- Grok Bot or Claude coordinates the agents.
What broke
I had set a 10-minute limit on voice calls. One parent was still talking when the limit hit, and the call just cut off mid-sentence. It felt really strange for them, and for me when I listened back. A hard cutoff is worse than no limit: if you cap calls, have the agent warn the caller and wrap up politely before the time runs out.
What changed along the way
At first, people were more careful talking to an AI. Now it feels normal to most of them.