The challenge
What was at stake
A B2B SaaS firm running outbound campaigns across 8 markets saw a 4.2% lead-to-booking rate, well below the 12% target.
Sales agents spent 78% of their time on voicemail drops and disqualified calls. The legacy IVR could not handle interruptions, struggled with regional accents, and suffered from 3-second response delays that killed conversion.
CRM synchronization was manual and stale, causing significant lead-follow-up delays of up to 8 hours.
The approach
What we built
We designed a parallelized agentic voice outreach system using a real-time WebSocket pipeline. 8 specialized AI agents (intent, dialogue, synthesis, CRM, escalation) managed the workflow.
The system handled 450 simultaneous outbound sessions, processed 55,000+ dials per week, and slashed end-to-end latency to 420ms. The system integrated directly with the CRM, pushing structured lead data in real-time, allowing agents to act on hot leads within 60 seconds of call completion.
Inside the system
Key capabilities
Real-Time WebSocket Pipeline
Streaming synthesis with 420ms latency, enabling natural interruptions and eliminating dead-air gaps.
Parallelised Outbound
Runs 450 simultaneous sessions across 8 markets, processing 55,000+ dials weekly with time-zone-aware scheduling.
Stateful Dialogue Controller
Adapts pitch and objection-handling sequences dynamically; reduced scripted prompt failures by 92%.
60-Second CRM Sync
Pushes structured lead scores and conversation summaries to Salesforce instantly, eliminating the 8-hour manual lag.
Under the hood
Technology
- Fine-tuned Whisper STT
- ElevenLabs-compatible TTS
- LiveKit WebSocket media server
- LangGraph state machine
- Salesforce REST API
The outcome
What changed
Lead-to-booking conversion rate increased from 4.2% to 11.8% (180% improvement). End-to-end latency reduced from 3,000ms to 420ms. Productive dial ratio increased from 22% to 84%.
- 4.2% → 11.8%
- Conversion rate
- 420ms
- Audio latency
- 500K+
- Calls processed