How Adyen does AI

Adyen says it has invested heavily in ML and AI talent and infrastructure, and that these systems already affect core business flows. Its approach combines real-time decisioning on every transaction, risk monitoring with transaction and KYC data, and newer GenAI work for operational support.

Reading notes

#

• Every transaction is processed with ML inference in real time on Adyen’s self-hosted big data platform.
• The company says it uses supervised, semi-supervised, and reinforcement learning to combat fraud, authenticate users, and improve payment completion.
• Adyen runs A/B/n testing with control groups and retrains or redeploys models on demand or on a weekly cadence.
• Its risk engine uses transaction and KYC data to monitor activity and detect behavioral patterns during onboarding and purchase.
• Detection methods include graph structures, graph neural networks, and deep neural networks.
• For checkout, fraud, authentication, authorisation, routing, and retry, the article describes a sequential ML setup with different models for different steps.
• Adyen says merchant rules can complement ML models, including policy rules and other rules for patterns not yet learned by the models.
• The company says it is researching ways to make decisions across the full funnel and combine fraud, conversion, and cost in one payment flow.
• It is also funding research with the University of Amsterdam’s AMLAB on causal inference and off-policy evaluation.
• Adyen describes its platform data as mostly structured and relational, with graph modeling used to capture deeper connections between entities.
• With banking licenses and financial services, the graph becomes larger and more directed, with more node types and more edges.
• The article says this leads to research on graph databases, sampling, graph features, and Graph Neural Networks for complex pattern detection.
• On GenAI, Adyen says it uses open-source foundational models hosted and fine-tuned in-house, and contributes back to the open-source community.
• GenAI work is focused on customer support and operations, especially ticket routing, summarisation, and context augmentation with RAG.
• The company says human-in-the-loop workflows are deliberate for privacy, security, and quality reasons.
• Adyen says LLM-based routing did not outperform a simpler TF IDF classifier in its use case.
• The GenAI stack is deployed on-prem on Adyen’s data clusters with GPUs, and is also made available internally to employees.
• The article closes by saying Adyen is exploring future topics such as hallucinations, synthetic data, self-alignment, and transformer-based architectures for mixed structured and unstructured data.