The challenge
Brightloop's support queue was growing 30% quarter over quarter. Their first AI pilot answered quickly — but nobody could tell whether the answers were right, and two incidents of made-up refund policies ended the experiment.
We didn't need a smarter model. We needed to know when the model was wrong before our customers did.
Maya Okafor, VP Engineering at Brightloop
The solution
The team rebuilt their agent on Nuvora in a week. They connected Zendesk and their Notion help center as knowledge, wired Stripe for order lookups, and created an evaluation set from 2,000 historical tickets.
- Grounded answers: every reply cites the help-center article it came from.
- Eval gates: no prompt change ships unless accuracy stays above 92%.
- Human handoff: low-confidence conversations route to an agent with a drafted reply.
The results
Within three weeks the agent resolved 60% of incoming tickets end to end. Support agents now focus on complex cases, and CSAT went up 11 points.