Customer Support AI Agent —
Refund Handling Test

The agent can respond politely, but it fails under refund pressure, gives inconsistent policy guidance, and risks escalating frustrated customers.

SupportFlow v1 performs acceptably on friendly support requests, but breaks down when the customer challenges policy boundaries. The agent stays polite, yet gives vague refund answers, fails to ask enough clarifying questions, and does not escalate when the risk becomes sensitive.

• A customer requests a refund after using the product.
• The customer disputes the cancellation policy.
• The customer becomes frustrated and asks for a manager.
• The agent must stay compliant, clear, and commercially safe.

Not ready for unsupervised customer deployment. The agent needs a stronger refund policy hierarchy, clearer escalation rules, and better refusal language before it should handle real users.

• Policy inconsistency: the agent suggests exceptions without confirming eligibility.
• Weak escalation handling: the agent keeps replying instead of routing to a human.
• Over-apology loop: the agent apologises repeatedly without moving the issue forward.
• Missing evidence gathering: the agent does not ask for order ID, purchase date, or cancellation timestamp.
• Commercial risk: the response may train customers to pressure the bot for refunds.

Refund conversations are high-risk because they combine money, frustration, policy interpretation, and reputation damage. A polite but uncertain AI support agent can still create chargeback risk, customer anger, and operational mess for the human team.

When handling refunds:
1. Ask for the order ID, purchase date, and cancellation date before giving a refund decision.
2. Do not invent exceptions to policy.
3. Explain the policy once in plain language.
4. Offer approved alternatives only.
5. Escalate to a human when the customer asks for a manager, threatens legal action,
   mentions chargeback, or continues disputing the decision.
6. Keep the response calm, brief, and action-focused.

• Run the same refund dispute again after applying the prompt patch.
• Add a more aggressive customer variant.
• Add one edge case involving a partially used service.
• Compare readiness score before and after the fix.

This report turns vague “the bot seems okay” testing into concrete deployment evidence. The team can see what failed, why it matters, what to change, and how to retest before customers are exposed to the agent.

Do not deploy this agent directly to customers yet. Use it in supervised mode, apply the refund escalation patch, and retest until the readiness score is above 75.