Voice AI · Lending OpsProduction · 8-week build

An AI Calling Agent that automated 78% of inbound loan qualification calls.

A US-based consumer lender was losing qualified borrowers to a 6-hour callback window. We shipped a multilingual voice agent on Twilio + GPT-4o Realtime that answers in 12 seconds, qualifies in 3 minutes, and books a human follow-up — 24/7.

PDF · 3 pages · 8-week build
78%
Calls automated
6h → 12s
Avg. response time
4.2×
Booked demos
−71%
Cost per call

The challenge

A 6-hour callback window was killing conversion.

Inbound leads spiked at night and weekends — exactly when the human team was offline. Conversion dropped 47% on after-hours leads. Hiring a 24/7 call center quoted at $480K/year before a single loan was booked.

The solution

A real-time voice agent with safe human handoff.

We designed a bilingual agent that picks up in under 12 seconds, runs a 14-intent qualification flow, books a follow-up in Calendly, and warm-transfers complex cases to on-call humans with full context preloaded.

Tech stack

What we built it on.

Voice & Telephony
  • Twilio Programmable Voice
  • LiveKit Agents
  • Deepgram Nova-3 STT
  • ElevenLabs Turbo v2.5
Reasoning & Orchestration
  • OpenAI GPT-4o Realtime
  • LangGraph state machine
  • Function calling
  • Guardrails AI
Knowledge & Memory
  • Pinecone vector store
  • Postgres + pgvector
  • Redis short-term memory
Backend & Integrations
  • Node.js + Fastify
  • HubSpot CRM
  • Calendly API
  • Twilio SMS fallback
Observability
  • Langfuse traces
  • Sentry
  • Grafana + Loki
  • Custom QA dashboard

Timeline

8 weeks from kickoff to 100% production traffic.

  1. Week 1
    Discovery & Call Audit

    Listened to 200+ recorded calls, mapped 14 intents, defined success metrics with the ops team.

  2. Week 2
    Conversation Design

    Built dialog flows in LangGraph, wrote persona + guardrail prompts, scripted 60 edge cases.

  3. Week 3–4
    Voice Pipeline MVP

    Wired Twilio ↔ LiveKit ↔ GPT-4o Realtime. Hit sub-800ms turn latency.

  4. Week 5
    CRM & Calendar Integration

    Two-way HubSpot sync, Calendly booking, SMS fallback for failed transfers.

  5. Week 6
    Shadow Mode Pilot

    Ran in parallel with human agents on 1,400 calls. QA dashboard caught 31 prompt regressions.

  6. Week 7
    Tuning & Localization

    Added Spanish + Hindi voices, tuned barge-in handling, deployed regional accent models.

  7. Week 8
    Production Rollout

    Phased traffic ramp 10% → 100% with automatic human handoff for complex cases.

Results after 90 days

Hard numbers, not vibes.

  • 78% of inbound calls fully resolved by the agent — zero human touch.
  • Average answer time: 12 seconds, down from 6 hours.
  • 4.2× more qualified demos booked, with 92% show-up rate.
  • −71% cost per qualified call vs. outsourced call center quote.
  • CSAT 4.6 / 5 across 3,800 surveyed callers.

Deliverables

What shipped.

  • Production-ready bilingual voice agent (EN/ES/HI)
  • LangGraph orchestration with 14 intents + safe handoff
  • HubSpot + Calendly + Twilio SMS integrations
  • QA dashboard with call replay and prompt regression tests
  • On-call playbook and 30-day support retainer
"It's the cheapest, most patient SDR we've ever hired — and it never sleeps. The thing paid for itself in week three."
VP Operations · US Consumer Lender (NDA)

Explore more work

Related case studies.

Want a calling agent like this?

We typically ship a production voice agent in 6–10 weeks. Free 30-minute scoping call.

Book a Strategy Call