Agent 16 · E-Commerce
ReturnsBot Agent
Decides returns by policy and flags serial returners.
Agent PackCommerceLangGraph
Found259
Fixed + verified251
Held for approval8
Recall1.00
One offline run on demo data with planted problems (recorded platform test, November 2026).
How ReturnsBot heals data, step by step
INData inSynthetic returns linked to Olist-shaped orders
01Detect259 problems found
02Diagnose5 problem types named
03Plan251 safe to fix · 8 need a person
04Fix + verify251 fixed and re-checked
05ReportPlain-English report and Control Tower record
- Client problem
- Return decisions are slow and inconsistent; abuse goes unnoticed.
- What it does
- Checks return window, order dates, photo evidence and customer history.
- How it heals
- Auto-approves clean returns, requests photos, rejects outside-window ones with a polite message.
- Human gate
- Serial returners and high-value refunds go to a human.
- Typical buyer
- D2C brands
- Built with
- LangGraph self-healing loop
- Demo data
- Synthetic returns linked to Olist-shaped orders (Olist is CC BY-NC-SA: synthetic for client demos).
- Tables
- returns ← synthetic
orders ← Kaggle olistbr/brazilian-ecommerce - Target outcome
- Target: 60% of returns decided automatically.
- Path to production
- Commerce / Support Desk Pack.
Issues detected on the demo data
| Issue | Risk | Proposed fix | Count |
|---|---|---|---|
| order_date_error Delivered before purchase – OMS clock or entry error | low | Flag order record for OMS correction | 3 |
| serial_returner Customer has 4 returns | high | Hold refund for fraud review | 8 |
| ready_to_approve Eligible return (wrong_item, day 1) | low | Approve and trigger refund | 117 |
| outside_window Requested 17 days after delivery (policy 10) | medium | Decline with policy explanation | 110 |
| photo_needed 'damaged' claim without photo | low | Ask customer for a photo | 21 |
How we demo it
- Load returns and orders.
- Run ReturnsBot.
- Show the approval queue.
Built on AWS
Runs in the platform container on EC2 Graviton (or on demand in Lambda), writes healed tables to the S3 lakehouse, uses Amazon Bedrock only to write explanations, and reports six metrics to CloudWatch.
Step into DigiQuench.
Agents that heal your data.
Watch an agent find planted problems, fix the safe ones and wait for your approval on the rest. Thirty minutes, live.