LLMs in Fraud Detection: Near-Term vs. Long-Term Roles

Context

You are designing fraud detection for a large-scale digital payments platform with:

• Real-time transaction scoring requiring sub-100 ms p95 latency and high availability.
• Multiple data sources: structured transaction/device/network signals; unstructured text (support chats, claims, merchant descriptions, KYC docs, emails); and semi-structured logs.
• A human review workflow for escalations and investigations.

Task

1. Compare large language models (LLMs) to traditional supervised models for fraud detection across:
   • Data modalities unlocked (text, logs, documents)
   • Feature engineering vs. representation learning
   • Accuracy (new pattern discovery vs. steady-state classification)
   • Latency and cost
   • Interpretability and governance
   • Robustness to adaptive adversaries
   • Privacy/compliance
   • Lifecycle operations (training, deployment, monitoring, updates)
2. Describe the near-term and long-term roles LLMs should play in this stack.
3. Propose a hybrid architecture that integrates LLMs and traditional models to balance accuracy, latency, cost, and risk.
4. Propose an evaluation plan (offline and online) to justify adoption, including metrics, ablations, cost/latency modeling, and risk controls.