Gradual Rollout: Test with real traffic, commit when you're confident

When a CX or AI team deploys a change, every user finds out at the same time whether it works. Even a fast rollback means thousands of real conversations already routed to a broken version. For high-volume deployments, that is not a testing strategy.

Gradual Rollout lets teams validate actionbook and model changes against live traffic before committing to a full deployment.

How it works

• Automatic candidate storage: Every version promoted from Dev to Staging is saved as a candidate automatically. No tagging, no manual steps.
• Two-slot architecture: Production runs a Live slot (current version) and an Idle slot (candidate under test). Conversations split between them by ratio.
• Configurable traffic split: Set the initial ratio from the dashboard using integers between 1 and 99. Start at 1%, increase as confidence grows.
• Conversation-level metrics: Every conversation is tagged Live or Idle. Resolution Rate, CSAT, and other core metrics are broken down by slot and by ratio segment.
• Promote or rollback in one action: Candidate wins - swap slots, Idle becomes Live at 100%. Something looks wrong - cut Idle to 0% immediately. The Live slot never changes.
• Model comparison support: Not limited to actionbook changes. Teams evaluating model updates can route live traffic to a candidate and compare performance before committing.