Top 3 that I think should be considered are these:

•Edge Cases and Hallucinations: Test how the agent handles tricky, out-of-scope, or vague questions to ensure it doesn’t give incorrect information.

•Integration and Data Accuracy: Verify that it correctly pulls information from your knowledge base and connects seamlessly with your help desk or CRM.

•Escalation Pathways: Check that the handoff to human agents is smooth, fast, and passes along the full conversation history.