Good point! For delegated outbound calls, I’d treat the evidence bundle as a separate, user-facing artifact rather than just call analytics. Analytics tells us how the call performed technically; the artifact tells the user what was actually achieved, ideally with supporting transcript excerpts, timestamps, and recordings where available.

Regarding Twilio, I’d use the answered callback (CallStatus=in-progress) as the answer boundary and completed as the call-end boundary. Neither guarantees success — voicemail or an IVR can answer too.

I’d argue that successful completion is an application-level outcome, determined by the agent’s workflow and backed by evidence, rather than a Twilio event.

Curious how you’re handling cases where the agent believes it got confirmation, but the conversation is ambiguous?

I’m the founder of Lucien, where we are testing that outbound pattern for one low-stakes business call at a time. I’m curious how you would model it in this architecture: as call analytics, or as a separate user-facing artifact—and which Twilio event would you treat as the authoritative boundary between connected, answered, and successfully completed? If you would like to pressure-test the current evidence bundle, the beta is linked from my name. Calling requires a card and prepaid credit, and Lucien is not for bulk marketing or emergency calls.