1) Why AI has become essential in the fight against fraud<\/strong><\/p>\n\n\n\n 2) Oneytrust’s AI toolkit<\/strong><\/p>\n\n\n\n 3) Humans + AI: augmented expertise, not automated<\/strong><\/p>\n\n\n\n 4) A strategy to acculturate our employees to Generative AI<\/strong><\/p>\n\n\n\n
The boom in language models and content generation tools has lowered the barriers to fraud: flawless phishing emails, carefully crafted fake profiles, synthetic identities that are more difficult to detect, industrialisation of fraud scenarios, reusable attack scripts. Static controls are no longer sufficient. We need systems that are capable of learning, detecting weak signals, and continuously evolving.<\/strong> AI is not a gadget: it is essential for detecting anomalies, linking scattered events, and responding in real time without hindering legitimate processes.<\/p>\n\n\n\n
Over the years, Oneytrust has built a range of scores to analyse digital events, identify fraud patterns and prioritise actions. Our scores aggregate more than 200 signals and are based on a shared database covering more than 15 million consumers, authorised by the CNIL (French Data Protection Authority) in 2013 \u2014 well before the GDPR.<\/strong>
Beyond fraud at a given moment, we observe velocity: how a user, payment method, device (phone, PC, etc.) or email address behaves over time. This intelligent memory makes it possible to streamline the journey of good customers (controlled reuse of acquired trust) and tighten the noose on risks (accumulation of clues, inconsistencies, anomalies). Our rule engines remain adaptive: we add, adjust and remove rules as trends evolve, to keep pace with changing attacks without compromising the customer experience.<\/p>\n\n\n\n
AI identifies, humans understand.<\/strong> Our investigators and fraud experts play a particularly crucial role in challenging alerts, contextualising cases, and revealing emerging patterns that models have not yet learned (low volumes, ambiguous signals, attacker innovations). This feedback loop feeds the models, improves the rules, and ensures an explainable decision. In particularly sensitive situations, human arbitration protects the customer relationship, avoids false positives, and maintains trust.<\/p>\n\n\n\n
As soon as ChatGPT was released in 2022, we quickly decided to train our teams in the use of generative AI to increase speed, operational efficiency and quality on non-sensitive tasks<\/strong>: writing, summarising, analysis assistance, preparing deliverables, etc. As part of the BPCE Group, we have access to the secure MAIA portal, which allows our employees to access reference models (e.g. GPT-4o, Mistral, Gemini) in a safe and ethical environment. In our training courses and with the help of our Generative AI Ambassadors, we emphasise the importance of maintaining a critical perspective on the results generated by AI (systematic verification, non-disclosure of sensitive information, traceability of uses): the aim is to accelerate useful production, not to automate indiscriminately.<\/p>\n\n\n\n