Every call passes a gate. Every word is checked afterwards.
An AI agent that says the wrong thing to a real person is a regulatory problem, not a bug report. Warmline is the layer that decides whether a call may be placed at all, then verifies what the agent said — disclosure, approved claims, commitments, opt-outs, market-sensitive news — and writes the result back as data.
Talk to the agent
A real voice agent, not a script. You play the prospect. When the conversation ends, the transcript ElevenLabs stored is run through the same post-call checks as every other call on this page — so if you can talk it into inventing a price or ignoring an opt-out, you will see it caught.
Walk a call through the policy layer
Three questions, in order: should we call this person, did the agent say anything it should not have, and what happened. A call that fails the first never reaches the second.
Try to get a claim past the checker
Type anything as though the agent had said it. This is the deterministic tier that runs on every transcript — no model, same code as production. It will not catch everything, and it says so: a sentence it cannot place is reported as unclassified rather than waved through.
The checker works on text, not audio. The same rules would apply to an SMS, WhatsApp message or email — and there they can run before the message is sent, which is a stronger control than flagging a call after it has ended.
Type something, or pick an example.
Prospects
Whether each record is callable at this moment, and the last call on it. Five of the six are seeded to fail a different check, so the policy layer is visible without reading the tests.
| Prospect | Number | Can we call now? | Last call on record |
|---|