Claude Main
Highest-scoring candidate after filtering the current preview data by this workflow.
90Workflow guide
Safest AI Agent for Refund Policy Replies
Evaluate which agents avoid unsafe refund or credit promises while still writing helpful support replies.
Best for: Support leaders, compliance reviewers, and customer-experience teams
Mistral Main
Prioritizes critical-failure rate, then score.
Doubao Main
Prioritizes cost tier, then workflow score.
Which agent best respects refund and policy boundaries?
This page does not replace human review. It reframes the leaderboard around a concrete buying and launch question. Before production, review raw outputs, business boundaries, and model versions.
Relevant task evidence
| Chinese Customer Complaint Triage | Qwen Main | 85 |
| Chinese Invoice Dispute Reply | OpenAI Main | 85 |
| Refund Policy Boundary Reply | OpenAI Main | 96 |
| English Security Questionnaire Answer | OpenAI Main | 96 |
| Japanese Appointment Intent Classification | Claude Main | 92 |
| Japanese Support Escalation Note | Claude Main | 92 |
| Spanish Support Reply for Wrong Item | Claude Main | 89 |
| Spanish Billing Cancellation Reply | Claude Main | 91 |
Failure tags to watch
literal_translation: 39unsupported_claim: 32unsafe_refund_promise: 29weak_cta: 22missing_field: 19too_verbose: 17
Average score: 80