TLDR: The CASE (Conversational Agent for Scam Elucidation) framework, developed by Google, utilizes Agentic AI and Gemini LLMs to combat social engineering scams in digital payments. It features a Conversational Agent that interviews potential victims to gather detailed scam intelligence and an Information Extractor Agent that converts these narratives into structured, actionable data. Deployed on Google Pay India, this system has increased scam enforcement recall by 21% and improved response times, demonstrating a scalable and safe method to bridge the intelligence gap created by off-platform scam orchestrations.
The rapid expansion of digital payment platforms has brought unparalleled convenience to global commerce, but it has also created fertile ground for sophisticated social engineering scams. These scams often begin and are orchestrated on platforms outside of the payment system itself, such as social media or messaging apps, making it incredibly difficult for payment platforms to detect and prevent them using only transaction-based signals.
To address this critical intelligence gap, researchers at Google have developed CASE (Conversational Agent for Scam Elucidation), a novel Agentic AI framework. This framework is designed to collect and manage user scam feedback in a safe and scalable manner, providing a deeper understanding of scam methodologies and patterns.
How CASE Works: A Dual-Agent Approach
The CASE framework operates with two core components, both powered by Google’s Gemini family of Large Language Models (LLMs), specifically Gemini 2.0 Flash for production deployment:
1. The Conversational Agent: This user-facing component is uniquely designed to proactively interview potential victims. Unlike typical chatbots that answer questions, this agent asks dynamic, investigative questions to elicit detailed intelligence about a scam. It adopts the persona of a specialist fraud analyst and is guided by strict interaction guidelines to ensure safety, empathy, and privacy. For instance, it never promises refunds or offers financial advice. The agent’s primary goal is to capture key facets of the scam, such as the initial contact method, the ‘hook’ used to build trust, and the specific action that led to financial loss.
2. The Information Extractor Agent: Once a conversation with the Conversational Agent is complete, the unstructured transcript is passed to this backend system. The Information Extractor processes this raw conversational data and converts it into a structured, machine-readable format, typically JSON, according to a predefined data schema. This crucial step transforms qualitative insights into actionable data for downstream automated analysis and manual enforcement mechanisms. It can accurately differentiate between scam and non-scam reports and classify the type of scam (e.g., fake loan, fake jobs).
Seamless System Flow and Robust Safety
The system’s operational flow is integrated within the user’s existing in-app support interface. It involves a real-time Intelligence Collection Phase where user input is simultaneously processed by a Safety Filter LLM and the Generator LLM. A decision logic module then determines the final response to the user, ensuring adherence to safety policies. The complete conversation is stored, and in an asynchronous Data Processing Phase, the Information Extractor processes these transcripts in batches.
Given the sensitive nature of financial scam discussions, CASE incorporates a multi-layered safety architecture. This includes inherent safeguards within the Gemini models, a dedicated Input Safety Filter LLM, and guided prompt architecture with negative constraints to prevent harmful outputs like unauthorized financial advice.
Real-World Impact and Results
The CASE framework was implemented and evaluated within the Google Pay India ecosystem. The results have been highly promising:
- Enhanced Scam Detection: By augmenting existing anti-abuse features with the new intelligence gathered by CASE, Google Pay India observed a significant 21% uplift in the volume of scam enforcements.
- Improved Enforcement Velocity: The availability of structured intelligence from user reports led to a substantial reduction in the time required to take action against malicious actors.
- High User Engagement: The Conversational Agent proved effective at maintaining productive dialogues, with over 45% of users who initiated a session answering three or more follow-up questions, indicating meaningful, in-depth interviews.
- Strong Safety and Quality: Pre-launch evaluations showed 99.9% compliance on egregious violation policies and 99.2% on sensitive topic policies. Post-launch, these figures remained high, with no egregious violations recorded in real-world interactions.
- Accurate Information Extraction: The Information Extractor achieved 83.8% accuracy in differentiating between scam and non-scam reports and 75.1% accuracy in classifying the specific type of scam.
The structured data generated by CASE is actively integrated into the anti-abuse ecosystem, supporting both manual analysis by human reviewers for complex cases and automated enforcement through machine learning models that detect and prevent new scam patterns at scale.
Also Read:
- Navigating the Landscape of AI Agents: Methods and Real-World Applications
- Lowe Lintas India and Google Gemini Collaborate to Revolutionize Daily Life with AI
A Generalizable Blueprint for Trust & Safety
While initially focused on Google Pay India, the architectural design of CASE is highly generalizable. It offers a reusable blueprint for building similar AI-driven intelligence systems in other payment platforms and can be adapted to broader Trust & Safety domains, such as combating online harassment, hate speech, and misinformation. The flexibility of the LLM system allows for adaptation to new contexts through strategic prompt customization and domain-specific examples.
Future work for CASE includes incorporating multimodal inputs (like audio recordings or screenshots), further automating evaluation and enforcement, and exploring ecosystem-level collaboration to share anonymized scam intelligence with external enforcement agencies or financial institutions. You can read the full research paper here.
