Voice AI
Building Voice AI That Customers Actually Don’t Mind Talking To
Sub-second latency, honest escalation and no infinite loops — the design choices that separate tolerable voice AI from infuriating.
Most voice AI deployments fail for the same reason: they are designed to avoid transferring to a human. The system is optimised for containment when it should be optimised for resolution. Customers do not mind a bot that handles a simple booking quickly and transfers the rest. They mind a bot that traps them in a loop.
Hear how an automated assistant sets expectations and hands a caller to staff in our Voice AI staff transfer demo.
Latency is a feature, not a tuning parameter
A turn-taking delay over 800ms reads as “the machine is thinking about me.” Over 1.5s it reads as broken. Budget your pipeline so transcription, model inference, and speech synthesis fit inside that window, and measure it on the 95th percentile, not the average. Averages hide the callers who hang up.
Design the escalation before the conversation
Decide upfront which intents transfer immediately, which transfer after one clarification, and which the agent owns end to end. Write the transfer prompts so the human who picks up knows what the caller wants and what the bot already tried. A warm transfer with context costs you nothing and saves the caller from repeating themselves.
Never let the bot apologise in a loop
The single most maddening voice AI behaviour is the apology loop — “I’m sorry, I didn’t catch that” repeated until the caller gives up. Cap misunderstandings at two, then transfer. A graceful exit is worth more than a stubborn attempt at containment.
Get these three things right and callers will use the system, thank the agent who handles the hard cases, and never mention the bot. That is the goal: invisible when it works, honest when it cannot.
