Explain the accountable refund-review loop
Separate the risk queue, human decision, and historical dashboard to answer the original automation question.
Question for this chapter
What is automated, who owns the decision, and what is not yet persisted?
Confirm the dashboard's scope
The refund_operations dashboard summarizes the ten existing refund_history rows. It does not show
the twelve risk-assessment rows or the Agent result from your session.
- Total Refund Requests (statistic) — 10 requests
- Total Refund Amount (statistic) — $3,444.91
- Decision Distribution (donut) — 7 approved, 1 partial, 1 pending, 1 rejected
- Refund by Reason (bar) — 4 defective, 2 changed_mind, 2 not_as_described, and one each for the rest
- Refund Amount by Customer (bar, Top 10) — historical refund total per customer
- Recent Refund Requests (data table) — up to 100 rows ordered by request time descending
Complete the solution
| Evidence | Confirmed result | Operational role |
|---|---|---|
| Batch risk assessment | 5 medium, 7 low, 0 high | Human review order |
| Ontology relationships | 10 customers, 12 orders, 6 connected refunds | Context for one request |
| Agent runtime demo | fixed order response, policy RAG, human confirmation | Learn the recommendation-decision flow |
| Historical dashboard | 10 past refunds totaling $3,444.91 | Existing operations view |
State the final decision
QuickReturn reviews five
mediumorders first and sevenloworders in the standard queue. There are nohighorders in this seed. Evidence enrichment, risk triage, policy retrieval, and the AI recommendation are automated, but a person confirms every final decision. Downstream Actors can return results after confirmation; real refunds, emails, and audit persistence remain outside this implementation.
Completion check
- Enriched 12 orders with customer and refund history.
- Confirmed the 5, 7, and 0 risk distribution and used it to order review.
- Traced one customer-order-refund path.
- Distinguished policy RAG from the fixed order-tool response.
- Confirmed the boundary between the AI recommendation and human decision.
- Can distinguish Actor results from persisted records.