Back Market's fraud team builds AI detection system in one week, contributing to €1.2M savings initiative
- Industries
- B2B SaaSMarketplace
- Company Size
- 201-1000
- Departments
- ITOperationsData
Key Highlights
- Around €100K in estimated fraud prevented in 5 months through AI-powered claims analysis
- Part of broader fraud prevention initiative projected to save €1.2M annually
- Built in one week with Dust by the fraud team without engineering resources
- Less than one day to adapt to new fraud tactics, compared to months previously
About Back Market
Back Market is Europe's leading online marketplace for refurbished electronics, connecting consumers with certified refurbished devices from trusted sellers. Operating across multiple countries, the company processes thousands of transactions daily, making fraud detection and prevention critical to protecting both the business and maintaining customer trust.
Challenge: scaling fraud detection while maintaining customer trust
Logistics fraud is a persistent challenge for e-commerce platforms. At Back Market, fraudsters target high-value electronics by purchasing expensive items, requesting refunds, and either sending back empty boxes or manipulating shipping labels so packages never reach their destination.
The stakes were high. Back Market estimated that 0.3% of all parcels were potential fraud cases, representing significant GMV loss. But implementing stricter verification processes for all refunds wasn't the answer.
"The issue is that it's not a simple problem," explained Aurélien Gervasi, AI Engineering Manager at Back Market. "You cannot put into place a complex and deep verification process for each refund request because it can create a very high level of strong detractors."
A few years earlier, facing elevated fraud levels, the company implemented a complex refund verification process that triggered public backlash. Customers complained loudly about how difficult it had become to get legitimate refunds.
The company built an expert fraud team capable of identifying fraud signals while maintaining a customer-friendly experience. But the team was constrained by their tools. Manual investigations relied heavily on SQL queries, and advanced capabilities required requesting engineering resources to build custom solutions or machine learning models.
"The investigation process takes several hours to sometimes a few days," said Gervasi. "As you can imagine, it's not something that scales easily with Back Market's business."
Solution: the fraud orchestrator Agent
Back Market launched a comprehensive fraud prevention initiative targeting logistics-related losses. As part of this broader effort, the fraud team identified an opportunity to leverage Dust AI agents. They envisioned a "Fraud Orchestrator," a master AI agent that would mimic the investigative process of their best fraud analysts, rapidly processing data-intensive fraud detection tasks.
The fraud team built a multi-agent architecture. The Fraud Orchestrator serves as the central coordinator, routing work to specialized sub-agents. Each expert agent focuses on checking a specific element of potential fraud cases:
- The Address Check agent evaluates delivery address risk by comparing against known fraud addresses.
- The Return Distance agent calculates geographic distance between delivery and return locations, identifying suspicious cross-border anomalies.
- The Customer Search agent analyzes customer history, calculating order count, lifetime GMV, and incident frequency to spot high-risk patterns.
- Additional specialized agents check for Tracking Incidents, Payment Incidents, and perhaps most innovatively, Conversation Patterns.
The Conversation Pattern agent represented a breakthrough capability. The fraud team had discovered that fraudsters often use similar opening messages when filing claims. They wanted to build a repository of these patterns and automatically compare new claims against them.
"If you try to do it manually, it's very time-consuming of course," Gervasi noted. "And if you try to do it with SQL, it's also not something that works very well."
When the fraud team identifies a new fraudulent message template, they simply update a Confluence page, and the system immediately incorporates it. No engineering team involvement required.
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 drop-off location in Ireland, and flagged that the customer had a 50% refund rate—an extremely strong fraud signal.
"This is why generative AI is very good for this," said Gervasi. "It has this ability to not only look at numbers but also extract insights from it and provide some explainability behind that."
The fraud team built this system themselves in roughly one week. When they need to adapt to new fraud tactics, updates take less than one day, sometimes just a few hours.
"The fraud team was able to build that autonomously," Gervasi emphasized. "They are fully autonomous on that."
Results: meaningful contribution to fraud prevention, with team autonomy as the breakthrough
Back Market's comprehensive fraud prevention initiative is projected to save more than €1.2 million annually. While it's difficult to isolate the precise impact of individual components, the team estimates that the AI-powered claims text analysis alone has prevented nearly €100,000 in fraud over five months.
But for Gervasi, who leads AI transformation at Back Market, the financial impact tells only part of the story. The more significant outcome is how the fraud team gained autonomy and velocity.
"The very interesting thing for me as an AI transformation manager is that they were able to do this autonomously," he said. "They were able to build this solution very quickly. It took them roughly one week of continuous work, and today when they need to update it, it takes them less than one day."
The fraud domain evolves constantly. Fraudsters adapt their tactics and develop new message templates. Previously, responding required going through engineering teams with limited bandwidth. Now, the fraud experts adapt their AI agents directly, iterating as quickly as the fraudsters themselves evolve.
Gervasi sees the Fraud Orchestrator as a template for how domain experts across the organization can become AI builders.
"If you think about your organization, you have a lot of expert teams that have this kind of specific use cases," he noted. "This is a clear example of how Dust enables these domain experts to become AI builders. They are able to solve their own problems with generative AI without relying on engineering resources, which are always a little bit restricted in terms of bandwidth."
At Back Market, that shift has transformed how the company approaches scaling fraud detection. Rather than building more complex rules or hiring more investigators, they've empowered their existing experts to work at a higher level, focusing on complex edge cases and strategic fraud prevention while AI agents handle the data-intensive investigative groundwork.
The result is a fraud detection capability that scales with the business, adapts to emerging threats in days instead of months, and contributes meaningfully to fraud prevention while maintaining the customer experience that Back Market's reputation depends on.
Interested in learning more about how Dust can help your team? Visit our solutions page or reach out to our sales team.
