EVA-Bench
Evaluate whether a voice agent completes an enterprise task accurately while sustaining a natural spoken conversation.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
Tara Bogavelli and coauthors describe 213 scenarios covering airline support, healthcare HR and enterprise IT service management. ServiceNow publishes the implementation.
Methodology
Automated callers hold complete spoken conversations with the tested agent. Separate EVA-A and EVA-X metrics assess task accuracy and conversational experience. The protocol validates the simulated caller and supports repeated trials, accents and background noise. Automated simulation and model judges are part of the measurement. Preserve caller, judge, audio condition and trial settings when comparing results. Paper revision v3 is dated September 9, 2026; it is not a benchmark edition.