Voice AI Agent Call Counter: 13 Exam Scenarios and Cost Down to the Penny

An engineer set up a voice AI agent for a small manufacturing company's phone line. The system uses Voximplant for telephony, Yandex Realtime (speech-realtime-250923) for speech, and Claude for creating order cards. Initially, the system had a 4-second delay, but by using a bridge server, the first sound comes within 554-804 ms. The engineer then built a custom metrics and cost tracking system, logging 27 fields per call into SQLite. Three key insights emerged: personal and spam calls were initially misclassified as target calls, awaited coroutine cancellation could silently drop call logs, and models often ask for phone numbers they already know. The engineer also found that token costs dominate (66% of expenses), and that a cheaper Yandex model performed worse. After verification, the counter's accuracy was within 1% of billing. The cost per target call is 6.8-8.4 rubles, and the engineer created 13 exam scenarios to test the system. The article concludes that API-based metering is more honest than subscriptions for reselling services.