- Home
- Case Studies
- AI Call Centre Case Study
AI Call Centre Case Study
Voice-first AI agent handling real inbound calls over SIP- built, tested, and production-ready.
The Challenge
Most AI voice products on the market start from the wrong end. A chatbot is built, a text-to-speech layer is added, and the result is an experience that sounds like a robot reading from a script, with the awkward pauses, unnatural cadence, and rigid responses that everyone has learned to endure or hang up on.
We believed a genuinely useful AI voice agent had to be designed from the telephone inward. That meant starting with real SIP telephony and real voice engineering, and then building the AI intelligence on top of that solid foundation. We had 14 years of telecom voice experience. We had access to the best LLMs and speech processing tools available. We built the platform to prove the architecture works and to be deployable for the first enterprise client who needs it. That client has now arrived: the platform is running in a live pilot with a mobile network operator in Africa, answering real subscriber calls on the operator’s own network.
Platform Status: Built, tested, and production-ready. This is an architecture case study documenting what we built and how- not a proof-of-concept.
What We Built
AI voice agent on the operator’s network
- The agent sits on the operator’s SIP trunk behind the existing customer care short code, handling calls on the live network rather than a test line
Subscriber self-service journeys
- Balance and usage enquiries, data bundle and plan information, plan changes, service activations, and first-line complaints
Account-aware conversations
- The agent identifies the subscriber from their calling number (MSISDN) and retrieves their profile before the first word is spoken
Operator knowledge base
- Operator tariffs, bundles, policies, and FAQs loaded into a RAG knowledge base so every answer is grounded in the operator’s current offer
Warm handoff to the existing contact centre
- Low-confidence or out-of-scope calls are transferred to the operator’s human agents with a live conversation summary
Operator console
- Live call volumes, containment and escalation rates, intent mix, and full transcripts for the operator’s care and QA teams
Outcome
- The pilot moved the platform from a proven architecture to a proven deployment: real subscribers, real network conditions, and real call volumes on an operator’s live care line.
Technology Stack
The platform is built as independent, swappable components. Each layer can be replaced (a different LLM, STT engine, or cloud) without re-engineering the rest.
Telephony & Voice
- Asterisk
- SIP trunking, Voice Activity Detection (VAD)
Speech-to-Text (STT)
- OpenAI Whisper (primary), Google STT (alternative)
LLMs
- OpenAI GPT-4o
- Anthropic Claude
- Qwen Omini
Text-to-Speech (TTS)
- Orpheus 3B
- Qwen Omini
RAG & Knowledge
- LangChain
- pgvector (PostgreSQL)
Integrations
- Operator SIP core / care short code
- Operator CRM and billing APIs
- Custom REST APIs
Frontend (FE)
- Angular
- TypeScript
Backend (BE)
- Node.js
- Python (FastAPI)