Inbound Tier-1 deflection

The primary entry point for Nortinia AI Call Center on inbound traffic is Tier-1 deflection. That means the voice agent replaces the old IVR ("press 1 for English") and resolves the most common call types on its own before any human gets involved. After six months the numbers are reasonably clear.

The death of the old IVR

Instead of a 4-8 second DTMF maze ("press 1, press 2, press 9 to go back"), the inbound call now looks like this: 1.2 seconds after the ring picks up, the voice agent says hello and asks an open question — "how can we help?" The customer states the problem in their own words.

While the STT (speech-to-text) is running, the engine is already classifying: within a median of 1.9 seconds the call is bucketed into one of 5-7 main intent categories. From there either the AI resolves it or it routes to the right human queue.

This is not the same as a button-driven IVR. One small SME customer hesitated for six months before switching, then sent us their CSAT three weeks later: old IVR 3.2/5, new AI 4.1/5. The single reason: customers no longer have to learn a menu map.

The four most-commonly deflected reasons

Based on live traffic, Tier-1 has four call types that make up 78% of all deflected calls:

1. Order / shipment status (31%) — "where is my package?" Database lookup by order number, instant answer. If there's a problem (delay, damage), routed to human.
2. Opening hours and contact info (22%) — "are you open tomorrow?" Purely static information, high-confidence reply.
3. Billing and payment questions (15%) — "did my payment arrive?" Database lookup. Disputes go to human.
4. Appointment scheduling and rescheduling (10%) — "can I move my booking to next week?" Calendar integration handles it automatically, unless the request has special conditions.

The remaining 22% spread across 30+ rare categories. Not worth writing a deflection rule for; they go to a human.

The numbers

Six months of live traffic across 14 customers, ~31,000 inbound calls:

• 42% resolved by AI alone — never reached a human
• 38% AI-to-human handoff — AI started it and passed context along
• 20% immediate human routing — sensitive or complex

Median length of deflected calls: 86 seconds. Median length of human-handled calls: 4 minutes 12 seconds. The gap isn't only because AI is faster (it is), but because deflectable calls are inherently simpler.

The handoff threshold: 2 unrecognized turns

The most important decision the AI makes on its own is when to escalate. The logic is simple and deliberately conservative:

• 2 consecutive turns where intent classification confidence is low → human queue.
• Emotional signal (raised tone, repetition, frustration keywords) → immediate human.
• Explicit complaint or grievance → automatically human, AI doesn't even try.
• Customer asks ("I want to speak to a human") → instant, no questions.

These four rules account for the 20% immediate handoff. We deliberately decided to hand off too often rather than risk a bad AI experience. The false-handoff rate (where AI escalated unnecessarily) sits at 8%, which is high but an acceptable trade-off.

What we never deflect

There are a few call types that never touch the AI:

• Emergencies ("my house has been broken into") — straight to human from second one
• Cancellations or contract termination — retention value mandates human
• Complaints or legal disputes — always human, also for legal reasons
• VIP customer tier (tenant-defined tag) — fixed human pool

These categories make up roughly 6% of calls, and AI doesn't attempt them even at high classification confidence.

Rollout time

For a typical SME, Tier-1 deflection takes 3-4 weeks to roll out:

• Week 1 — voice-sample collection, FAQ ingestion, intent map
• Week 2 — Telnyx SIP integration, parallel running (AI listens, human answers)
• Week 3 — AI takes 20% of calls, tuning
• Week 4 — full live, 40%+ deflection rate

Those four weeks sound more painful than they are. Most of the time is not the technical integration but the FAQ organisation and the intent map. Until the SME can clearly state what questions get what answers, the AI can't decide on their behalf either.