Initial research article published September 26, 2026.
What it does
Feedzai markets a fraud and financial-crime platform spanning transaction monitoring, scams, account opening, payment fraud and case operations. The technical value proposition is real-time scoring from transaction, customer, device, behavioral and network signals, combining machine-learning models with rules and investigator workflows. It is not a single model and implementation outcomes depend on available labels, integration latency and the operating policy around scores.
Actual ML versus marketing
Machine learning can rank risk, detect nonlinear interactions and adapt models more efficiently than static rules. It cannot independently determine legal liability, customer intent or the proper friction threshold. Feedzai public case studies report outcomes such as a 37% improvement in detection for one customer and alert-backlog reduction for another; these are vendor-selected claims. They do not disclose enough common baseline, loss maturation, false-positive cost or independent replication to generalize.
Evidence gates
| Gate | Minimum evidence | Why it matters |
|---|---|---|
| Baseline | Current model, fraud rate, approval and review policy | A lift percentage needs a denominator |
| Labels | Fraud, scam, first-party, dispute and recovery definitions | Different labels imply different controls |
| Performance | Recall, precision, false positives, dollars saved and friction | AUC alone does not establish operating value |
| Maturity | Observation window and aged losses by cohort | Early results can miss delayed chargebacks |
| Governance | Reason codes, versions, overrides, fairness and monitoring | Challenge, audit and controlled change |
| Resilience | Latency, capacity, fallbacks, replay and incident response | Real-time rails leave little recovery time |
| Economics | License, integration, review cost and prevented loss | Adoption does not establish positive ROI |
Implementation burden
A mature deployment requires event schemas, historical labels, streaming integration, identity linkage, case-management design, sanctions/BSA boundaries, model validation and feedback loops. Human-in-the-loop review should have explicit authority, service levels and escalation. Champion/challenger testing should segment by channel, merchant, geography, payment rail and customer tenure; aggregate gains can hide a deteriorating niche.