BNP Paribas: real-time fraud detection and faster KYC
A governed foundation first, then the models that run on it.
Fraud caught as it happens, KYC cut by more than half, on a data estate where nothing may leave the perimeter it belongs to.
Project Context
BNP Paribas works on a massive and sensitive data estate. Two problems sat on top of it, and both came back to the same cause.
Fraud has to be caught while the transaction is still moving, on volumes no manual review can follow. And KYC, the check that opens or blocks a customer relationship, was slowed down by customer data scattered across systems and by controls done by hand.
Neither is a modelling problem. A detection model trained on a fragmented customer reference finds fraud that is not there and misses the fraud that is. The foundation had to be fixed first.
Strategic Objectives
Detect while it still matters
Fraud detected after settlement is fraud recorded, not fraud stopped. Detection had to run on the flows, at the volume the bank actually processes.
A customer known once, not five times
Customer data spread across systems means five partial versions of the same person. KYC cannot go faster until there is one reference to check against.
Regulatory data that holds
Quality was not measured, so nobody could say how reliable a regulatory figure was. What is not measured cannot be defended in front of a regulator.
Nothing outside the perimeter
The whole programme runs under heavy compliance and sovereignty constraints. Speed could not be bought by loosening either.
Solutions Delivered by Adservio
A governed data foundation first, then the models on top of it. Master data management gives a single customer reference, quality is measured continuously rather than audited occasionally, anomaly detection runs on the flows, and a data copilot answers the analysts with sources they can follow.
One customer reference through MDM
Master data management reconciles the scattered customer records into a single reference. That is what makes a KYC check a lookup rather than an investigation.
Quality measured continuously
Data quality stops being an occasional audit and becomes a measurement that runs. A regulatory figure now comes with a known level of reliability.
Anomaly detection on the flows
Detection models are wired into the transaction flows rather than run on extracts. Fraud is caught while the transaction is still in motion.
A data copilot for the analysts
Analysts query the data in plain language and get answers that carry their sources. The decision goes faster without the audit trail getting thinner.
Results
Detection up by 40%, on flows rather than on extracts, which is what makes the difference between recording fraud and stopping it.
KYC cut by 60%, once a single customer reference replaced the reconciliation done by hand.
Model accuracy up by 30%: the same algorithms perform differently once the foundation underneath them is reliable.
Impact
Detection that acts instead of recording
Running on the flows moves detection from after the fact to during. A fraud caught in motion can be stopped; the same fraud found in a report can only be counted.
A relationship opened in hours, not days
Cutting KYC by 60% is not an internal efficiency gain. It is the time a customer waits before the bank can work with them.
The foundation as the real lever
Accuracy up 30% without changing the approach shows where the gain came from: not from better models, but from a customer reference and a measured quality underneath them.
More Client Work
INES programme: AI-augmented depositary control
AI FastPass & Mobile Order
A Workday migration mapping accelerated by AI
Augmented CIO: GenAI in the data software factory
Augmented CIO: GenAI at the heart of the SDLC
Augmented CIO: a sovereign AI foundation across the SDLC
Interbank flows & SAP S/4HANA modernisation
360° Performance & Observability
Dentaline: deep learning on panoramic X-rays
EBIOS RM risk analysis & GDPR
Application Performance Centre
DevOps automation for life insurance
Building an educational nonprofit information system
Cinké Évolution : data visualisation steering
EBIOS RM risk analysis on Azure
SIAG modernisation via SAP Fiori
DevOps & DataOps: Azure cloud migration
Redesign & digital maintenance: web performance
Fabrik Digitale: cloud-native on Azure
TIV Test & Reporting: rail maintenance
GOC 2.0 project: critical architecture audit
Digital transformation: Soon & Connect
SpeedBooking: a platform shipped for 600+ hotels
A time-management rules engine rebuilt with no regression
Fix the foundation, then the models
Let's talk about your customer reference, your data quality and your detection requirements. An Adservio expert gets back to you within 24 business hours.