How Kadrella's AI Calls a Real Person When It Needs a Decision

Most of the day-to-day running of Kadrella is handled by Claude, connected to the company's own API. Most of it does not need a human: rescheduling a lesson, chasing a late payment, answering a parent's question. A small number of situations do, and for those, Kadrella's AI does something unusual: it calls a real phone.

Why a chat message is not enough

A text notification is easy to miss and easier to ignore, especially for someone already deep in other work with notifications off. A ringing phone is different: it is the one signal almost everyone still answers, because for decades a phone call has meant "someone needs you right now."

What actually happens on the call

When Claude hits a situation that needs a human decision, a refund to approve, a student waiting with no teacher assigned, it places an actual WhatsApp voice call, says what is happening in a synthesized voice, listens to the spoken answer, and turns that answer back into an action. No app to open, no dashboard to check first.

• Claude decides a human is needed and describes the situation as text.
• The text is turned into speech and a WhatsApp call goes out.
• The spoken answer is transcribed and read back into the conversation with Claude.
• Claude acts on the answer: approves the refund, reassigns the lesson, or whatever the decision was.

Built for one company first

This is not a lab demo. It runs in production for Kadrella itself, handling the calls its own team gets when something needs a decision faster than a message would get read. The full engineering story, the WhatsApp voice stack, the speech pipeline, what broke on the way to something reliable enough to depend on, is written up in detail: How Claude Calls Me on WhatsApp to Manage My Company.

There is also a short video of the call itself: Claude is calling me on WhatsApp.

What this means if you run a tutoring business

The everyday admin, reminders, invoices, follow-ups, can run on its own. But automation that quietly fails without telling anyone is worse than no automation at all. The standard here is that anything Claude cannot resolve on its own gets escalated to an actual person, on an actual phone call, rather than silently dropped or buried in a notification nobody sees.