Review refunds with AI recommendations and human decisions

A refund request that asks about both speed and accountability

The fictional electronics retailer QuickReturn receives everything from small accessory returns to high-value laptop refunds. Having a person rediscover every fact is slow. Letting a risk score or an AI recommendation approve the request can miss policy and customer context.

Today, you are the QuickReturn support and risk specialist answering this question:

Which parts of refund review should be automated, and at what point must a person own the final decision?

Decision criteria for this mission

• Perspectives in play — Support operations automates repeated evidence gathering, risk review checks the score and policy basis, and a human reviewer owns the final decision and accountability.
• Decision constraints — The current synthetic seed contains 5 medium, 7 low, and 0 high cases. The Agent's fixed learning response and Actor outputs do not mean that a refund, email, or audit record has been persisted.
• Completion signals — Enrich the evidence for 12 orders, use risk to order the review queue, and explain why an AI recommendation grounded in policy cannot reach downstream action before human confirmation.

Workshop goals

• Enrich orders with customer and refund history in one review bundle.
• Use risk as a review-order signal, not as an approval outcome.
• Retrieve policy evidence before an AI recommendation.
• Separate the AI recommendation from the human's final decision.
• Distinguish a post-confirmation Actor result from a persisted business record.

Seven decisions

flowchart LR
    accTitle: Seven decisions that separate refund recommendations from human decisions
    accDescr: Start with a refund request, enrich its evidence, retrieve policy, require a human decision, check recommendation and record boundaries, and explain the accountability loop.
    request["1. Define the request"] --> evidence["2. Enrich evidence"]
    evidence --> risk["3. Order the queue"]
    risk --> policy["4. Retrieve policy"]
    policy --> human["5. Require human decision"]
    human --> check["6. Check accountability"]
    check --> explain["7. Explain accountability"]

Each chapter follows question → reason → action → observation → interpretation → next decision. The goal is not “let AI approve,” but “know what evidence is gathered and where automation stops.”

Before you begin

• A support-operations or risk-analyst account with Editor access or higher
• Enough download space for a scenario ZIP of about 37 KB
• To run the Agent, an environment where an administrator configured an API key for the imported claude_sonnet_4_6 model
• About 90 minutes

You do not need a terminal, Python, or a clone of dhub2-examples. Without model credentials, you can still complete the batch, graph, and dashboard work and inspect the Agent contracts in the Agent chapters.