Konnect's existing setup routed every inbound call through a static IVR into one of fourteen agent queues. Average handle time was creeping past nine minutes; the cost of adding capacity was linear in headcount. They wanted AI in the loop — but with real telephony, real CRM lookups, real handoff. Not a chatbot wearing a phone.
We rebuilt the call path around a routing brain that owns the conversation until escalation. Telnyx handles SIP and recording; OpenAI Realtime handles the voice loop; an internal policy engine decides when a human takes over.
Two-week sprints. Live staging from week three. We shadowed the existing call centre in week six, ran 5% live traffic in week ten, and went to 100% in week fourteen. The cutover was reversible the entire way.
Three engineers from TIN. Two voice ops leads from Konnect. Friday demos every sprint, no surprises.
Average handle time dropped 68% in the first month and held. Customer CSAT moved from 7.4 to 8.1. Konnect cancelled a planned hiring round of 22 L1 agents and redeployed budget into expansion markets.
The architecture has been live for fourteen months at the time of writing. We retain a small ongoing pod for new policy work and Realtime model upgrades.
Node + Hono · Postgres + pgvector · Redis · Telnyx · OpenAI Realtime · Salesforce · Terraform on AWS · Datadog. SOC 2 Type II controls. PCI-DSS in-scope handling for billing transfers.