GLOSSARY
Evaluation
An evaluation assesses an AI system using defined tasks, criteria and methods to examine its capabilities, limitations and risks for a particular use.
Test what matters in daily sales work
A negotiation assistant should do more than produce fluent responses. It may need to distinguish agreed deadlines from proposals and keep contradictory information visible. An evaluation turns these requirements into testable cases. Include ordinary work and rare situations with serious consequences. Scoring needs explicit criteria: a reply may be courteous while containing an unauthorised commitment. One overall score could conceal that difference. Unfinished or abandoned tasks also belong in the assessment rather than disappearing from the reported results.
Measure performance under comparable conditions
Compare quality, working time, costs and checking effort against a baseline. Qualified reviewers should distinguish error types and explain disputed assessments. A benchmark can support initial selection; internal tasks reveal operational suitability. After a model, source or workflow changes, check whether previous capabilities remain intact. A useful evaluation therefore delivers more than a successful demonstration. It identifies the conditions under which the system works, when people must intervene and which failures should prevent approval for the intended use.