Northwind SaaS · B2B Software
A support copilot that deflected 40% of tickets
We built a retrieval-augmented support copilot grounded in the client's docs and ticket history, with evals gating every release.
- 40% ticket deflection
- 3.2s median response
- 92% answer-accuracy on evals
The challenge
Northwind's support team was drowning in repetitive questions that were already answered somewhere in their documentation. Off-the-shelf chatbots hallucinated and eroded trust.
What we built
A retrieval-augmented generation (RAG) system grounded strictly in Northwind's documentation and resolved-ticket history, wrapped in a support copilot embedded in their help center.
- A hybrid retrieval pipeline over docs and historical tickets
- Strict grounding with citations, plus refusal when confidence is low
- An evaluation harness scoring accuracy and groundedness on every deploy
Results
Within six weeks the copilot was deflecting 40% of inbound tickets while maintaining a 92% accuracy score on the eval set — and crucially, the team trusted it because every answer carried a citation.