BCG X built the FAQ agent for one of the largest telecom operators in the world. Kejue is the platform it runs on, inside the operator's own data centre.
- Industry
- Telecommunications
- Location
- On the operator's own hardware
- Calls
- Inbound
- Languages
- Arabic and English
BCG X was engaged by one of the largest telecom operators in the world to put a voice agent on the main question line. BCG X designed and built the agent. Kejue supplied the platform it was built on, and deployed that platform inside the operator's own data centre.
The line takes the questions every operator takes. Billing, coverage, devices, plans, roaming, what a charge was for, what a customer is entitled to. The answers live across a large and often contradictory body of internal material, and an answer that is close but wrong is worse than no answer at all. Nothing about those calls was allowed to leave the building.
BCG X wrote the prompt, the guardrails and the knowledge grounding, and graded the agent on a set of real questions scored for whether the answer was right rather than whether it sounded right. The agent answered 85% of them accurately, and the customers it served rated the experience at 89%.
What the phone line
looked like.
- A question line carrying the volume of one of the world's largest subscriber bases.
- Answers spread across a large body of internal material, some of it out of date and some of it contradictory.
- No call audio, transcript or customer record could leave the operator's own infrastructure.
- A plausible wrong answer costs more than a transfer, so accuracy had to be measured, not asserted.
The platform
it was built on.
The platform, inside their building
Speech recognition, synthesis, the agent runtime and every transcript run on the operator's own hardware, with GPU nodes for the models. Kejue deployed and supports it there. Nothing crosses the perimeter.
BCG X wrote the rules, the platform enforces them
Answer only from the knowledge base. Never state a figure that is not in the retrieved passage. Say it is a virtual agent when asked. BCG X authored each rule; the platform applies it in the call path rather than leaving it to the prompt.
When it does not know, it says so
The call that cannot be answered from the knowledge base ends in a handover, not a guess. Transcript, the search the agent ran and the latency on each turn are on the record, so a disputed answer can be traced rather than argued about.
Graded, then watched
BCG X scored the agent against a set of real questions before launch, and the same measures stay on a board the operator and BCG X both read afterwards, with alerts when either moves.
The numbers
BCG X reports.
- 85%
- answer accuracy across the graded question set
- 89%
- customer satisfaction on the calls it handled
- On-prem
- the whole platform inside the operator's own data centre
