The questions I get asked most often about AI in a support or fintech operation, answered directly — with the conditions and trade-offs still attached, because an answer without them is not usable.
Direct answers, with the caveats left in.
Build the one where a human already does the task every day, the correct answer is easy to recognise, and a wrong answer is cheap to undo. Those three conditions are what let you finish the project: the workflow gives you data, the recognisable answer gives you a test, and the cheap undo gives you permission to ship.
Map every escalation rule a human currently applies, and automate only the work that sits above the rules. Escalation is not a feature you add to an automated flow — it is the boundary of the flow, so the boundary has to exist and be written down before anything is automated. Teams that automate first and define escalation later end up with a system that handles everything confidently, including the things it must not touch.
A voice agent is production-ready when it can end a call safely without ever needing to be clever. That means four things are tested against real audio, not transcripts: it recognises the caller's actual speech, it handles being interrupted, it has somewhere to hand off to a human, and every action it takes against a real system has a limit it cannot exceed.
Evaluate the vendor's failure behaviour, not their accuracy claim. Ask for the worst ten calls they can show you, what the system did when it was wrong, and who found out. A vendor who can answer that has built something; a vendor who answers with a benchmark number has built a demo.