AdvancedAgentic AIChallenge #6

Refund Decision Agent

Build an agent that reasons over customer refund requests, weighs the policy, and autonomously decides the outcome — while knowing exactly when to hand a tricky case to a human.

Challenge Overview

This challenge drops you into the returns desk of an e-commerce store, where an agentic refunds bot clears a queue of 6 refund requests— each with its own context, reason, and customer history.

For every request the agent reasons over the refund policy and picks the outcome — Full Refund, Partial Refund, Replace, or Deny— records the justification, and judges whether the case is risky enough for a human escalation. The skill on test is confident decision-making on the clear cases, paired with knowing the limits of autonomy on the rest.

Decision Policy

Evaluate in order — the first matching rule wins.

1.Item defective / damaged on arrival

Replace

2.Wrong item shipped

Replace

3.Outside 30-day window (and not defective)

Deny

4.Changed mind · unused / sealed · in window

Full Refund

5.Changed mind · opened / used · in window

Partial Refund

When to Escalate

Escalation is independent of the decision — you still pick the action, but flag it for a human if either signal is present.

Order value above $500

Escalate to human

Customer has 3+ prior returns (abuse signal)

Escalate to human

Step-by-step Instructions

1. Open the test system and read the first refund request and its context

2. Check condition, reason, days since purchase, value, and prior returns

3. Apply the decision policy in order — the first match is your action

4. Pick the justification reason that matches your chosen action

5. Decide whether to escalate to a human (high value or repeat returns)

6. Submit and continue through all 6 requests for your score

What you will learn

✓ Multi-factor reasoning over a policy with priority ordering

✓ Choosing between several action paths, not just yes / no

✓ Human-in-the-loop design — knowing the limits of autonomy

✓ Detecting abuse and risk signals before acting

✓ Justification / audit-trail thinking for every decision

✓ Real tool exposure: UiPath Agents, Power Automate + AI Builder, Automation Anywhere AI Agent Studio, Blue Prism decision flows

Common mistakes

• Refunding a request that is past the 30-day window

• Giving a full refund on an opened, used item

• Auto-approving a high-value order instead of escalating

• Missing an abuse pattern (3+ prior returns)

• Denying a defective item instead of replacing it

Complexity

Advanced · Agentic

Policy reasoning, multiple action paths, and human-in-the-loop judgment.

Scoring

Correct decision10 pts
Correct reason5 pts
Escalation (2 cases)5 pts each
Maximum score100 pts
Start Challenge