Case Study: Voice Customer Service Agent
Phone and in-app voice support with tools, safety, handoff and real-time SLOs
Model interview answer for a phone and in-app voice support agent (20,000 calls a day, about 30 percent of turns calling a backend tool, refunds as the irreversible action): a text-versus-voice difference table, a cascaded architecture with per-stage p50 and p95 budgets and a Monte Carlo of voice-to-voice latency including tool-call turns (without a spoken filler 39.4 percent of turns breach 1.5 s of silence and 26.0 percent breach 2 s; with a filler 13.6 and 1.6 percent), narrow tools with a read-back refund state machine and a simulation of the residual wrong-refund probability (6.90 percent with no confirmation, 0.475 percent with read-back and a spoken yes, 0.208 percent with active confirmation, and why an inattentive yes dominates ASR yes/no errors), caller authentication and audio prompt-injection controls, a handoff context packet, an evaluation harness with simulated callers, bootstrap intervals, shadow-mode agreement and the 2,629 calls per arm needed to detect a 3-point success change, a back-of-envelope cost model against a human agent with break-even containment between 28 and 36 percent, a staged rollout with binomial kill-switch rules, and hedged compliance notes on consent, PCI and retention. All numbers come from fixed-seed scripts with assumed rates and prices; none is a measurement of a real ASR, LLM, TTS or telephony system.
Practice questions (6)
-
View →
Designing Confirmation for a Voice Refund
Advanced · Free -
View →
Allocating a Voice-to-Voice Latency SLO
Advanced · Free -
View →
Human Handoff and the Context Packet
Advanced -
View →
Designing the Evaluation Harness and Shadow Phase
Advanced -
View →
Cost per Call versus a Human Agent
Advanced