Compliance monitoring · Production

Back Market's fraud team builds AI detection system in one week, contributing to €1.2M savings initiative

The problem

Back Market faced persistent logistics fraud where fraudsters purchased high-value electronics, requested refunds, and returned empty boxes or manipulated shipping labels. The investigation process took several hours to days, advanced capabilities required engineering resources, and an estimated 0.3% of all parcels were potential fraud cases representing significant GMV loss.

First attempt

A previous complex refund verification process triggered public backlash as customers complained about difficulty getting legitimate refunds. Manual SQL-based investigations could not scale, and SQL was ineffective for conversation pattern matching.

Workflow diagram · grounded in source
1
Fraud claim submitted
trigger
“purchasing expensive items, requesting refunds, and either sending back empty boxes or manipulating shipping labels so packages never reach their destination”
2
Orchestrator routes to sub-agents
routing
“The Fraud Orchestrator serves as the central coordinator, routing work to specialized sub-agents.”
3
Address risk check
ai_action
“The Address Check agent evaluates delivery address risk by comparing against known fraud addresses.”
4
Return distance analysis
ai_action
“The Return Distance agent calculates geographic distance between delivery and return locations, identifying suspicious cross-border anomalies.”
5
Customer history analysis
ai_action
“The Customer Search agent analyzes customer history, calculating order count, lifetime GMV, and incident frequency to spot high-risk patterns.”
6
Conversation pattern matching
ai_action
“They wanted to build a repository of these patterns and automatically compare new claims against them. When the fraud team identifies a new fraudulent message template, they simply update a Confluence page, and the system immediately inc…”
7
Structured risk output
output
“When the Fraud Orchestrator analyzes a case, it provides structured output for each check: a risk level and an explanation. In one example, the agent identified that the delivery address in Spain differed dramatically from the return dro…”
8
Fraud team adapts patterns
feedback_loop
“When they need to adapt to new fraud tactics, updates take less than one day, sometimes just a few hours.”
Reported outcome

The AI-powered fraud detection system contributed to a fraud prevention initiative projected to save more than €1.2 million annually, with AI claims analysis alone preventing nearly €100,000 in fraud over five months.
The fraud team can now adapt to new fraud tactics in less than one day and operates fully autonomously without engineering resources.

Reported metrics
fraud prevented through AI claims analysisnearly €100,000
Annual fraud prevention savings (broader initiative)more than €1.2 million
time to build AI detection systemroughly one week
Time to adapt to new fraud tacticsless than one day
Show all 5 reported metrics
fraud prevented through AI claims analysisnearly €100,000
annual fraud prevention savings (broader initiative)more than €1.2 million
time to build AI detection systemroughly one week
time to adapt to new fraud tacticsless than one day
estimated fraud parcel rate0.3%
Reported stack
DustConfluence
Source
https://dust.tt/customers/back-markets-fraud-team-builds-ai-detection-system-in-one-week-contributing
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The AI-powered fraud detection system contributed to a fraud prevention initiative projected to save more than €1.2 million annually, with AI claims analysis alone preventing nearly €100,000 in fraud over five months.

What tools did this team use?

Dust, Confluence.

What results were reported?

fraud prevented through AI claims analysis: nearly €100,000; Annual fraud prevention savings (broader initiative): more than €1.2 million; time to build AI detection system: roughly one week; Time to adapt to new fraud tactics: less than one day (source-reported, not independently verified).

What failed first in this deployment?

A previous complex refund verification process triggered public backlash as customers complained about difficulty getting legitimate refunds.

How is this compliance monitoring AI workflow structured?

Fraud claim submitted → Orchestrator routes to sub-agents → Address risk check → Return distance analysis → Customer history analysis → Conversation pattern matching → Structured risk output → Fraud team adapts patterns.

Related compliance monitoring cases