How to Test an AI Voice Agent Before Go-Live
How to test an AI voice agent before go-live: four layers, what each one cannot catch, and the run count you need to be 95% sure a 1-in-20 fault would show up.
How to test an AI voice agent before go-live: four layers, what each one cannot catch, and the run count you need to be 95% sure a 1-in-20 fault would show up.
Most failed AI voice agent transfers never reach the carrier: 11 of Vapi’s 16 named symptoms are config on your side. Diagnose and fix warm transfers.
A full-cycle AI sales agent owns all seven stages, from target selection to re-entry, with no human starting or finishing any of them. It is a coverage claim.
Retell AI publishes 20 concurrent calls by default and $0.10/min to Australian destinations. Synthflow gates campaigns to Enterprise and scopes by contract.
An employer using an AI agent to screen candidates is the deployer, not the provider. Article 26 requires human oversight, input control and six months of logs.
A Scam Likely label names the company to approach: it is T-Mobile’s, applied by First Orion. AT&T uses Hiya; Verizon and Spectrum use TNS. Triage costs nothing.
Minutes times rate is not the formula vendors bill on. Four published mechanics sit in between, including Retell’s prompt-token multiplier and Twilio rounding.
Mostly no, but voicemail counts: Twilio treats a call answered by voicemail as completed and charges it. Retell publishes a 10-second minimum on some calls.
Prompt engineering is never the long pole; carrier registration is. Twilio puts A2P 10DLC reviews at 10-15 days, so the SMS leg costs weeks, not days.
Three fields carry the answer: Vapi’s endedReason, Retell’s disconnection_reason and Twilio’s call status. Vapi’s spec lists 597 values against Retell’s 34.