Ninety-two percent correct

The team asks whether that is good enough to launch. You have not seen the errors or the evaluation dataset.

Consider task type, error severity, reversibility, review, baseline, and how errors are distributed.

Eight percent weak brainstorming suggestions may be manageable. Eight percent incorrect account actions may be unacceptable. Ask what “correct” means, whether cases represent real use, and whether a small critical category fails disproportionately.

Compare with a simpler baseline and evaluate human review or constrained scope. Launch readiness depends on consequences and controls, not accuracy in isolation.

Read the error distribution before deciding what users may rely on.