Breaking the scam cycle: AI intelligence redefines digital payment safety

Digital payment fraud is spiraling into a global crisis, costing users and platforms billions each year as scams grow more sophisticated and harder to detect. To address the rising threat of digital payment fraud, a team of researchers has introduced a scalable AI-driven system that strengthens fraud detection capabilities in real time.

The study, “CASE: An Agentic AI Framework for Enhancing Scam Intelligence in Digital Payments,” published on arXiv, details the deployment of an advanced conversational AI framework on Google Pay India to address the intelligence gap in detecting social engineering scams.

Understanding the scam detection challenge

The explosive growth of digital payments globally, particularly in markets like India, has been accompanied by a surge in social engineering scams that exploit unsuspecting users. Traditional fraud detection systems rely heavily on transactional and on-platform signals. However, these approaches often fail against scams that originate on external channels such as messaging platforms or social media, leaving significant blind spots in real-time fraud detection.

The research identifies this intelligence gap as a critical hurdle. While fraudulent transactions are executed within payment ecosystems, the manipulation often begins elsewhere, making it challenging for systems to flag malicious activity promptly. Manual interviews with victims have historically been an effective method of intelligence gathering, but these are labor-intensive and impossible to scale across millions of daily transactions.

To address this, the authors developed CASE (Conversational Agent for Scam Elucidation), a two-part AI framework that uses large language models (LLMs) from Google’s Gemini family. The system leverages advanced natural language processing to engage users in structured conversations, extract actionable insights, and feed this intelligence back into enforcement pipelines.

How CASE enhances fraud intelligence

The CASE framework operates through two tightly integrated components: a Conversational Agent and an Information Extractor Agent. Together, they create a seamless process for collecting, processing, and utilizing actionable intelligence against scams.

The Conversational Agent, powered by Gemini 2.0 Flash, is engineered to proactively conduct interviews with users who suspect fraudulent activity. Unlike traditional chatbots that only react to user queries, this agent initiates guided conversations, asking context-specific questions to build a comprehensive picture of the scam. It uses a robust safety framework that includes layered filters to detect and block unsafe content, strict guidelines to prevent the dissemination of financial or legal advice, and a privacy-first design that avoids processing personally identifiable information.

Once the conversation is complete, the transcript is processed by the Information Extractor Agent, which converts unstructured text into structured, machine-readable data. Using a schema-driven approach, it labels key data points such as whether the user was scammed, the type of scam involved, the platform of origin, and a concise summary of the incident. This structured intelligence is stored securely and fed into both manual review processes and automated enforcement systems to strengthen fraud detection and prevention.

By transforming fragmented, unstructured user narratives into structured intelligence, CASE empowers analysts and machine learning models to better detect emerging fraud patterns and respond more swiftly to evolving threats.

Evaluating impact and future potential

The deployment of CASE within Google Pay India has delivered measurable results, validating its effectiveness in live environments. In the months following the partial rollout, the system achieved a 21% increase in scam detection recall, significantly boosting the platform’s ability to identify and prevent fraudulent transactions.

Rigorous evaluations underpinned this deployment. Pre-launch testing involved adversarial red teaming and structured evaluations, ensuring the framework met stringent safety and performance standards. The Conversational Agent achieved 99.9% compliance with policies against harmful content and 99.2% compliance in handling context-sensitive topics. Human evaluators also confirmed high-quality interactions, with conversations consistently maintaining a respectful, empathetic tone while staying focused on eliciting critical details from users.

The Information Extractor Agent demonstrated strong accuracy in classifying scam-related data, recording 83.8% accuracy in detecting scam reports and 75.1% accuracy in categorizing the type of scam involved. Importantly, user engagement metrics revealed that more than 45% of users actively participated in extended dialogues, providing detailed insights that enriched the platform’s intelligence capabilities.

Beyond enhancing detection rates, the structured intelligence generated by CASE has accelerated enforcement velocity, reducing the time between scam identification and action against malicious actors. By combining real-time conversational data collection with asynchronous data processing, the framework ensures timely and scalable responses to rapidly emerging threats.

Expanding the framework’s applications

While the initial deployment focused on Google Pay India, the architecture can be adapted to other markets and platforms with minimal modifications, making it a versatile solution for global payment ecosystems.

The potential applications extend beyond fraud detection in financial services. By adjusting prompts and training examples, CASE could be adapted to address a wide range of challenges in Trust and Safety domains, including the detection of online harassment, misinformation, and other forms of digital abuse. Its modular architecture, combined with robust safety guardrails, positions it as a flexible tool for high-stakes environments where accuracy, safety, and scalability are essential.

However, the researchers also acknowledge the limitations of the current implementation. The initial deployment is English-only, despite the underlying models supporting multiple languages. Future enhancements are expected to integrate multilingual capabilities to broaden accessibility in linguistically diverse markets. Additionally, while the system currently relies heavily on prompt engineering, future versions may explore fine-tuned models for improved handling of niche conversational nuances.

The authors provide a roadmap for further innovation. Planned enhancements include support for multimodal inputs, enabling users to share evidence such as screenshots or audio clips, and the automation of evaluation and enforcement processes to improve scalability and efficiency. Another key priority is fostering industry-wide collaboration by sharing anonymized scam intelligence across platforms and with law enforcement agencies, enabling a more unified approach to combating digital fraud.