Latency is the product
Under roughly 800 milliseconds of response latency, callers behave as if they are speaking to a person. Above about two seconds, they start talking over the agent and the conversation degrades irrecoverably. Model choice matters far less than the streaming architecture around it.
Disclose immediately
Callers forgive an AI agent; they do not forgive being deceived about one. A single clear sentence at the start of the call measurably reduces hostility and abandonment.
Interruption handling and escalation
Barge-in support, graceful handling of silence, and a hard escalation rule after two failed intent resolutions. Define which intents are never handled by the agent, in writing, before launch — clinical advice, complaints and payment disputes usually belong on that list.
Review real calls weekly
The gap between test scripts and live behaviour is where deployments fail. A weekly review of a sample of transcripts, with prompt and routing changes shipped from it, is the difference between an agent that improves and one that decays.