CASE: An Agentic AI Framework for Enhancing Scam Intelligence in Digital Payments

1. Problem Statement

The rapid growth of digital payment platforms (specifically India's Unified Payments Interface, UPI) has been accompanied by a surge in sophisticated social engineering scams.

• The Intelligence Gap: Unlike traditional fraud, these scams are often orchestrated on external surfaces (e.g., messaging apps, social media) rather than within the payment platform itself. Consequently, the transaction data on the payment platform appears authentic, leaving anti-abuse systems without the necessary context to detect the manipulation.
• Limitations of Current Reporting: Traditional user reporting mechanisms (static surveys, free-text boxes) fail to capture nuanced details or allow for dynamic follow-up questions, resulting in ambiguous and non-actionable reports.
• Scalability Issues: While manual interviews by human analysts are effective, they are too expensive and slow to scale for millions of users.

2. Methodology: The CASE Framework

The authors propose CASE (Conversational Agent for Scam Elucidation), a novel Agentic AI framework built on Google's Gemini 2.0 Flash models. The system is designed to proactively interview potential victims and convert unstructured narratives into structured, actionable intelligence.

A. System Architecture

The framework operates in two distinct phases:

1. Real-Time Intelligence Collection (Conversational Agent):
   • Role: Acts as a specialist fraud analyst to conduct empathetic, investigative interviews.
   • Mechanism: Uses a Generator LLM guided by a sophisticated prompt architecture (Role Definition, Interaction Guidelines, Success Criteria, Privacy by Design). It asks dynamic follow-up questions to uncover the scam's Modus Operandi (MO).
   • Safety: A parallel Safety Filter LLM runs alongside the generator to assess user inputs for policy violations (e.g., hate speech, dangerous content) and ensure the agent does not provide financial advice or promise refunds.
2. Asynchronous Data Processing (Information Extractor Agent):
   • Role: Converts the raw conversation transcripts into structured, machine-readable data.
   • Mechanism: Uses In-Context Learning (ICL) with a schema-guided prompt to populate a predefined JSON schema.
   • Output: Extracts key fields such as is_user_scammed (boolean), possible_scam_mo (classification), scam_origin_surface, and a conversation_summary.

B. Evaluation Methodology

The system employs a rigorous, multi-layered evaluation framework adhering to Responsible AI principles:

• Safety: Tested via Structured Evals (adversarial datasets) and Red Teaming (human experts simulating bad actors). Policies cover high-severity violations and contextually sensitive topics (e.g., financial advice).
• Quality & Utility: Human expert raters evaluate Topic Adherence (staying on track), User Respect (empathy/termination handling), and Information Elicitation (success in capturing the scam MO).
• Accuracy: The Information Extractor is benchmarked against a "golden dataset" of manually curated transcripts to measure classification accuracy.
• Monitoring: A hybrid post-production approach uses automated evaluators (autoraters) calibrated against a statistically significant sample of human reviews.

3. Key Contributions

• Novel Agentic Pattern: Introduces a Dual-Agent Architecture (Interviewer + Extractor) specifically designed for high-stakes, emotionally sensitive investigative interviews, moving beyond standard transactional chatbots.
• Scalable Narrative Intelligence: Successfully bridges the intelligence gap by automating the collection of off-platform scam details, transforming unstructured user stories into structured data signals.
• Responsible AI Implementation: Demonstrates a robust safety framework for financial domains, including multi-layered guardrails, strict privacy-by-design (no PII passed to LLMs), and comprehensive human-in-the-loop evaluation.
• Generalizable Blueprint: Provides a reusable architecture for other Trust & Safety domains (e.g., harassment, misinformation) and payment platforms.

4. Results and Impact

The framework was deployed on Google Pay (GPay) India. Key metrics include:

• User Engagement: Over 45% of sessions involved users answering three or more follow-up questions, indicating the agent successfully maintained productive, in-depth dialogues.
• Safety Performance:
  • Pre-launch: 99.9% compliance on high-severity policies.
  • Production: 100% compliance on high-severity violations and 99.5% on contextually sensitive topics across ~3,600 samples.
• Extraction Accuracy:
  • 83.8% accuracy on binary scam detection (is_user_scammed).
  • 75.1% accuracy on multi-class Modus Operandi classification.
• Enforcement Impact: The integration of this new structured intelligence into existing anti-abuse systems resulted in a 21% uplift in the volume of scam enforcements.

5. Significance

The paper validates that augmenting traditional transaction-based signals with structured narrative intelligence is a highly effective strategy against social engineering scams.

• Strategic Value: It shifts the paradigm from reactive, signal-based detection to proactive, intelligence-driven prevention.
• Operational Efficiency: It solves the scalability bottleneck of human interviews, allowing platforms to process vast amounts of user feedback automatically while maintaining high safety standards.
• Future Outlook: The framework lays the groundwork for ecosystem-level collaboration, where anonymized scam patterns could be shared across financial institutions to create a coordinated defense against evolving online threats. Future work includes multimodal inputs (screenshots/audio) and expanded multilingual support.