Overview & Challenge
400 Agents, 90,000 Tickets A Month
Cedar Support Group's agents spent most of their day on the same twenty questions: password resets, refund status, plan changes, order tracking. Hiring couldn't keep up with growth, and wait times were climbing past eleven minutes.
The catch: support answers live inside order systems, not documents. A chatbot quoting help articles would have failed on day one — Nova had to do things, safely, inside real production systems.
Approach & Solution
Actions First, Answers Second
We built Nova as an action-taking agent, not a Q&A bot: verified identity, looked up the order, then executed the fix through audited tool calls — refund caps, dual approval above thresholds, and instant human handoff on low confidence.
- 22 audited toolsRefunds, resets, plan changes — every call logged and reversible.
- Confidence-gated handoffAnything uncertain routes to an agent with full context attached.
- Weekly quality reviewsAgents grade samples; corrections retrain the agent monthly.
Related Service
This project combined Machine Learning with Robotic Automation for safe, audited actions.
