The demo always goes well. The real test is the call the agent wasn't built for, and what it does next.
Most voice-agent demos show the happy path. A caller asks for an appointment, the agent finds a slot, everyone hangs up pleased. That part is genuinely easy now. What separates an agent you can leave answering your phones from one you can't is everything that happens off that path.
Decide the exits before the script
Before writing a single prompt, list the moments the agent should stop: a caller who is upset, a question about money it can't verify, anything medical or legal, anyone who simply asks for a person. Give every exit a destination, whether that's a live transfer, a callback slot or a message to a named inbox. "I don't know" should never mean a dead end.
Say less, confirm more
A good agent repeats back the details that matter, such as the date, the name and the callback number, and asks before it acts. It costs a few seconds per call and saves the calls that would otherwise go wrong silently.
Make the handoff warm
When the agent passes a call on, the person picking up should already have a summary: who called, what they wanted, what the agent tried. Nobody should have to ask the caller to start again.
Measure the misses, not the minutes
Average call length tells you very little. These lists tell you what to fix next week:
- Calls the agent transferred, and the reason it gave
- Questions it couldn't answer, word for word
- Callers who hung up partway through
