An application can treat Jev’s answer and its confidence as separate inputs. The answer says what the model selected; confidence helps decide whether to act on that selection. TypeSafe documents this as a confidence-gated routing pattern. The pattern is valuable only if the uncertain path has a real destination and the threshold is tested on the work you actually receive.
Begin with suggestions
For a customer-support queue, start by showing a proposed destination beside the original ticket. Have reviewers accept or correct it. Collect a small, representative set of these decisions, including ambiguous tickets and cases that carry a high cost of error. This phase reveals category overlap, missing context, and questions that sound clear but do not match your operational policy.
Evaluate the tradeoff
For several thresholds, count incorrect automatic routes and the number of tickets that would still need review. A higher threshold may reduce bad automatic decisions while sending more work to people. The right setting depends on the harm from a mistake and the team’s capacity; a number copied from another application is not a validation. Revisit the threshold when your traffic, question wording, or model version changes.
Keep failures visible
An API timeout is not “low confidence.” It is a call failure and needs its own handling. An empty or malformed state is an input problem. An unauthorized action must be blocked regardless of the model’s output. Record which path each ticket took so a reviewer can distinguish model uncertainty from a system failure.
For an implementation checklist, see Use confidence to route work to human review. The official confidence documentation explains how TypeSafe presents confidence alongside probabilities.
Published 24 September 2026. Pattern description checked against TypeSafe documentation on that date.