Agentic AI in Production: Lessons from 10M+ Voice Calls

Most agentic AI demos fall over the moment they meet a real customer. A demo has one happy path. A production voicebot handling BFSI and GovTech traffic across India, MENA, and SEA has thousands of paths, half of them created by people trying to break the flow on purpose.

What actually breaks

At LiaPlus, scaling from a first release to 10M+ calls exposed the same failure classes over and over:

• Ambiguous intent resolution. A single-agent chatbot can shrug and ask a clarifying question. A voicebot on a live call has maybe 400ms before the silence feels broken. Agentic routing has to guess well the first time.
• State drift across handoffs. The moment you split a flow across multiple agents (intent, KYC, retrieval, resolution), state gets lost at every boundary unless you're deliberate about what each agent is allowed to forget.
• Escalation as a first-class outcome. Autonomous decisioning on complex queries only works if "hand this to a human" is a designed exit, not a fallback bolted on after launch.

Why agentic beats IVR here

Traditional IVR trees are cheap and predictable but brittle — they only handle the paths someone thought to draw. Agentic workflows (used in VaniSetu and VaniPatra for government voice navigation and form-filling) let the system reason about intent instead of matching keywords, which is the difference between "press 2 for billing" and actually understanding what a citizen is asking for.

The tradeoff is observability. IVR failures are visible in the tree. Agent failures hide in reasoning traces, so you need to invest early in logging every agent decision, not just the final transcript.

What I'd tell a team starting today

Ship the escalation path before you ship the clever routing. Instrument every agent boundary before you need to debug it. And budget real engineering time for the 400ms latency target — agentic reasoning is naturally slower than pattern matching, and users notice the gap immediately on a live call.