Confidence score
A confidence score is the number that decides whether an AI agent sends its answer automatically or hands the ticket to a human. Get it wrong and the agent either escalates everything (useless) or auto-sends things it should not (dangerous).
Why the raw model score is not enough
Language models are systematically miscalibrated. Their raw certainty does not reliably track whether they are correct, so an answer that "seems sure" can still be wrong. Using the model's own vibe as the gate is how bots over-close tickets they should have escalated.
Confidence as a calibrated composite
Resolve computes confidence from several real signals, not one: retrieval strength (the strongest predictor of a correct answer), claim-level grounding, intent confidence, self-consistency, and the risk tier of the action. Those thresholds are calibrated against real reopen and CSAT outcomes, so the number means something. If any material claim is unsupported, the agent does not auto-send.
The tunable dial
Where you set the bar is a business decision, so Resolve exposes it as a plain-language confidence dial with a live impact preview. Raise it to auto-resolve only the safest tickets, lower it to let the AI handle more, and see the projected effect before you commit. No admin certification required.
Frequently asked
What signals go into a good confidence score?
Retrieval strength (the top predictor), claim-level grounding, intent confidence, self-consistency, and the risk tier of the action, all calibrated against real reopen and CSAT outcomes rather than the model's raw certainty.
Can I control how confident the AI must be before it auto-resolves?
Yes. Resolve exposes a plain-language confidence dial with a live impact preview, so you set the bar and see the projected effect before applying it.