Replacing Legacy IVR with an Agentic Voice Layer: A Case Study

The problem with classic IVR

"Press 1 for sales, press 2 for support" was designed for a world where compute was scarce and speech recognition was unreliable. Today both assumptions are wrong, but most enterprise contact centers still route on rigid decision trees. The cost shows up in three places: caller frustration, misrouted tickets, and agent handle time.

What "agentic voice" replaces it with

An agentic voice layer answers in natural language, asks one or two clarifying questions, and either resolves the call itself or routes it with full structured context to the right human team. Three components matter:

• Intent + entity extraction happens continuously, not at fixed menu nodes.
• Tool use lets the agent read the CRM, check order status, and open a ticket without escalation.
• Structured handoff — when a human is needed, they get the transcript, intent, sentiment and a recommended next action.

Case study: large enterprise telecom

A large enterprise telecom replaced a 12-year-old IVR serving millions of calls per year. The agentic layer was rolled out behind a feature flag, A/B-tested against the legacy IVR for 90 days, then promoted to the default.

Architecture

PSTN ─► SBC ─► self-managed telephony ─► media bridge
                                            │
                                            ├─► streaming STT
                                            ├─► intent router
                                            ├─► LLM agent (function calling)
                                            │     ├─► CRM lookup
                                            │     ├─► order system
                                            │     └─► ticketing
                                            └─► streaming TTS

What moved

• First-call resolution rose materially because the agent could read the CRM mid-call.
• Average handle time fell — human agents started conversations already briefed.
• Caller satisfaction improved most on the simplest calls (status checks, address updates), which the agent now resolves end-to-end.

Rollout pattern we recommend

1. Mirror traffic first — run the agent in shadow mode against real calls without speaking.
2. Promote one intent at a time (start with "check status"; end with "billing dispute").
3. Keep a one-key escape to a human agent for the first 6 months.
4. Treat the agent like any other production service: SLOs, alerts, runbooks.

What this is not

This is not "chatbot on a phone line." A chatbot answers; an agentic voice layer acts. The architectural difference — tool use, structured handoff, persistent context — is what makes the metrics move.