What Confidence Threshold for Auto-Resolution?
The confidence threshold for auto-resolution is the line above which an AI is allowed to send an answer on its own and below which it must hand the ticket to a human. There is no universal number, because the right threshold depends on ticket risk and must be calibrated against real reopen and CSAT outcomes. Critically, the score being thresholded should be a calibrated composite, not the model raw self-reported confidence, which is systematically miscalibrated.
Confidence should be a composite, not a vibe
A trustworthy confidence signal combines several inputs rather than trusting the model to grade itself:
- Retrieval strength: how well the knowledge base actually covers the question. The strongest predictor.
- Claim-level grounding: whether each material claim is supported by a source.
- Intent confidence: how sure the system is about what the customer is asking.
- Self-consistency: whether the model gives a stable answer.
- Action-risk tier: how costly it is to be wrong about this specific action.
Raise it for risk, lower it for coverage
Higher thresholds make the AI more cautious and increase escalations; lower thresholds increase automation and risk. Sensitive actions and high-value accounts should be held to a higher bar than routine questions. The goal is not the highest possible auto-resolution rate, it is the highest rate at which reopen rate and CSAT stay healthy.
How Resolve exposes the dial
Resolve turns this into a tunable confidence dial with a live impact preview, so moving it shows how auto-resolution and escalation change before you commit. Enterprise accounts are held to a higher bar automatically, and an angry ticket never auto-closes. You trade coverage for caution deliberately, in plain language, with no admin role or consultant required.