Voice agents fail without a spoken SOP

A voice bot without a hang-up policy is a hostage situation. Customers cannot see a spinner. They hear silence, then they shout, then they hang up angry. We write the spoken SOP before we pick a model: greet, confirm, escalate, close.

Latency is the product. Anything over about 800 milliseconds feels broken. We stream partials and never wait for a perfect sentence before we start talking back.

Budget the silence

The SOP names the maximum pause, the interruption policy, and the words we use when we did not hear the customer. 'Sorry, I missed that' is a designed line, not an improvisation.

We log every turn with timestamps. When a call goes badly, we replay the timeline, not a vibe about the model being 'off today'.

Escalate like a colleague

Warm transfer includes the transcript and the intent. The human should not ask for the account number again. If they do, the agent failed even if the model spoke fluently.

Voice is operations with a microphone. Write the procedure, then automate the edges.