Case study · enterprise-transform
Telecom Support Copilot
Grounded RAG assistant cut support deflection to 46% - evals before demos.
Telecom Support Copilot case studyOverview
The brief.
A telecom operator replaced ungrounded chatbots with a RAG support copilot grounded in its own knowledge base, deflecting 46% of tickets while keeping answers cited and eval-gated.
Client
Telecom operator
Telecom / Enterprise
millions of subscribers
Gallery
In the wild.
The numbers.
Results
46%
support deflection
+18%
CSAT
99%
cited answers
100%
eval gates in CI
The challenge
What was in the way.
Hallucination liability
Earlier bots confidently gave wrong plan and billing answers. In a regulated telecom, a wrong answer isn't just bad UX - it's a liability.
No evals
Every model or prompt change was a gamble; regressions were discovered by angry customers, not by tests. There was no way to know if it was actually better.
“Deflection hit 46% and CSAT went up, not down. The difference is every answer is cited - customers can verify.”
Outcome
The bottom line.
Grounded RAG assistant cut support deflection to 46% - evals before demos.
FAQ
About this engagement.
How do you prevent hallucinations?
Grounded RAG with enforced citations + confidence thresholds; low confidence falls back to a human.
What about evals?
A RAGAS eval suite gates every change in CI - no deploy on regression.
Is customer data safe?
Private infra by default; no public API calls without a DPA + masking.
Want results like these?
We build to the outcome, then let the numbers speak.