A three-phase programme for a B2B e-commerce operator: replacing fragmented rule-based fraud detection with ML scoring, unlocking high-value customer spend through intelligent credit limit modelling, then eliminating the residual review queue with a multi-agent system.
A major European B2B e-commerce operator was scaling transaction volumes across 12 markets. Fraud detection had been built country by country — each market running its own rule set, authored by whichever analyst or engineer happened to own the problem locally. There was no shared logic, no common standard, and no systematic way to measure whether any given rule was still doing useful work. The accumulated rule sets had grown large enough that monitoring their individual impact had become practically impossible.
Credit limit decisions were governed by the same rule-based approach — blunt thresholds that could not distinguish a genuinely high-risk order from a high-value one placed by a trusted customer. The result was a system that over-flagged legitimate transactions while providing no reliable protection against actual fraud.
The rule-based model was sending over 20% of all orders into manual review — a volume the operations team could not sustainably process. The backlog directly delayed order fulfilment and shipping, damaging the customer experience for what were, in most cases, entirely legitimate buyers.
The second problem was upstream of fraud altogether: the same conservative risk logic was setting credit limits too low for high-value customers. These buyers — the most commercially important segment — were being constrained from purchasing at the volumes they wanted, or abandoning the platform entirely. The business was losing revenue not to fraud, but to over-caution.
Two phases — operational bottleneck first, then revenue unlocking.
We replaced the fragmented country-by-country rule sets with a single ML-based fraud detection system, deployed consistently across all 12 markets. Orders are scored at the point of placement — high-confidence legitimate transactions clear automatically; genuinely anomalous ones reach the review team.
With the review backlog resolved, we addressed the revenue constraint. The rule-based credit model was replaced with one that could distinguish a high-value, low-risk customer from a genuinely elevated-risk order — something binary thresholds cannot do.
The residual manual review queue — orders the ML model flagged as uncertain — remained a human bottleneck. In Phase 3 we deployed a multi-agent system to handle this queue autonomously. Each flagged order is investigated by a coordinated set of AI agents that gather context, assess signals, and reach a decision, escalating to a human only when the case genuinely warrants it.
Let's talk about what's achievable for your specific context.
Start the Conversation →