What we learned deploying AI voice to 200 stores
Objection patterns, escalation triggers, and the one metric nobody warned us about.
The first 30 days of any new AI rollout reveal the things the spec sheet can’t.
Objections you can’t script
“How much under MSRP” comes in more variants than we modeled. The fix: hand callers to a human on price-talk if the AI’s confidence drops below 0.4.
The one metric nobody warned us about
Handoff latency. Customers waiting more than 1.2 seconds after “can I talk to a person?” bail at 3x the rate. We added a sentinel that fast-paths ANY user utterance containing the words “human”, “manager”, or “real person” — straight to your desk.
What’s surprised me is the good KPI: median call duration for fully-AI-handled calls dropped 22% — because the AI is unfailingly polite, doesn’t put the caller on hold, and asks one thing at a time.
If you’re rolling this out internally: time your escalation triggers, not your scripts.