Divert known fraud callers to an AI voice agent

Until now an automation rule had one action: block. The caller's number gets a SIP 403 on its next call, the call never reaches your route, and the fraudster moves on to the next carrier in their list. It is effective and cheap, and it tells the caller only that this trunk is closed.

Divert keeps the door open and changes who answers. The number is controlled the same way, but the next call is routed to a Voice AI destination of yours: an agent with a persona you chose on the agent side. The one we tested is an elderly person who is happy to chat, asks the caller to repeat things, and goes along with whatever the script wants, slowly. The fraudster spends their time on a machine. Your customer never hears the phone ring.

1. A call from the number is analyzed and the verdict crosses your rule's threshold. The number is added to Controlled numbers with the reason, the action and an expiry.
2. The next INVITE from that number reaches the ingress. Controls are checked before routing, so the divert happens at the SIP layer, before the call touches your route or your switch's downstream.
3. The proxy sends the call on to the Voice AI destination over TLS, with the agent's number as the request URI, and answers the agent's digest challenge if it has one. The media is relayed and recorded both ways, as always.
4. When the call ends, the recording is decoded (G.711 or G.722, whichever the agent chose), transcribed and analyzed. The record shows the call was diverted, and the transcript shows what the caller tried.

Nothing about the diverted call is special to the pipeline. It is billed per second like any other, it counts against your ports while it is up, and every piece of evidence is kept: the stereo recording, the transcript, the verdict, the SIP ladder, the capture, the quality of each leg.

1. Under Network, add a destination of kind Voice AI: the agent's SIP host, TLS on port 5061, and the agent's number as the Agent ID. For xAI that is sip.voice.x.ai and the Direct SIP number you registered with them. Digest credentials or an IP allow-list go on the agent side.
2. Under Automation, create a rule: the probability threshold, the recommendations that count, the action Divert the call, the Voice AI destination, and a duration.
3. Place a test call through the web dialer from a number you control, let the verdict cross the threshold, and call again. The second call is answered by the agent.

The Divert to a Voice AI destination is marked beta in the portal. The SIP side works. What is still settling is which agent providers we test against and how their personas are best written.

We do not know yet. Today's AI voices are good and getting better, but in a live call they still sound synthetic in places: the timing of a reply, the way a question is handled, the too-even tone. A practised fraudster who makes hundreds of calls a day may notice within a minute and hang up. If that is all that happens, the feature has still cost them a minute, kept your customer out of it, and left you a recording of the attempt with the number controlled for the duration you chose. That has value, but it falls short of a trap.

The SIP path, the controls, the recording and the analysis are in place and provider-agnostic. As voice agents get more natural, the same rule gets more useful without anything changing on your side. We will publish what we see on real traffic, including the calls where the caller caught on.

One practical note: a diverted call is a recorded call. Recording-consent rules vary by jurisdiction and apply here as they do to every call through the platform. Check yours before you point an agent at live callers.

A flagged number's next call can now be answered by a voice agent instead of your customer, at the SIP layer, with the whole call recorded and analyzed. It works today with xAI's voice agents. Whether the persona holds up against a real fraudster is still being tested. The divert, recording and analysis already work.

Related: three ways operators use Open Voice Shield, getting started, or start free with $5 credit.