Securing the Future: AT&T Uses Generative AI to Transform Fraud Protection

At AT&T, we are dedicated to providing innovative, reliable wireless carrier services to our 182 million wireless customers every day; we continually strive to improve operations, enhance customer satisfaction, and ensure security. Central to this commitment is our robust AI foundation, built on the Databricks Data Intelligence Platform, which we’ve used to modernize our approach to data and AI at scale.

This foundation has allowed us to address complex challenges, reduce inefficiencies, and implement scalable solutions across our operations. Now, we’re using it to explore how Generative AI (GenAI) can transform fraud prevention. Our first proof of concept (POC) highlights the exciting potential of this technology to proactively address one of the most critical issues in telecommunications: preventing sophisticated fraud attempts in real time.

From Traditional AI to GenAI: Elevating Fraud Detection

Fraudulent activities, such as account hijacking and identity theft, cost businesses and consumers billions annually and significantly impact customer trust. As fraudsters continue to innovate, it’s imperative that we stay ahead with adaptive, cutting-edge solutions. While traditional machine learning (ML) methods have been effective in identifying fraud patterns and generating risk scores, they often fall short in analyzing unstructured data, such as chat text, to identify additional aspects of fraudulent activity.

With traditional ML, we used models to analyze call transcripts (interaction between call center agents and customers), generate risk scores and flag potential fraud cases. These ML models helped us reduce fraud attacks by 80% with over 100 fraud detection ML models in production, improving customer protection while saving millions of dollars. However, fraud patterns evolve rapidly, and this reactive approach required manual interventions and extensive investigator effort.

Generative AI gives us a transformative advantage. We can easily identify complex patterns in unstructured data, recognize unknown fraud signals, and generate synthetic datasets to simulate real-world scenarios. We are proactively learning from all emerging fraud cases across different industries and applying these insights to our work. This allows us to stay ahead of potential threats and address issues before they escalate, ensuring robust protection and continuous improvement in our fraud detection efforts.

Fast and Easy Innovation with Databricks

The Databricks Data Intelligence Platform has been instrumental in accelerating our journey into GenAI. By leveraging a unified lakehouse architecture, we developed our fraud prevention POC using tools such as:

• Unity Catalog for secure, governed access to critical datasets.
• MLflow to manage the machine learning lifecycle, tracking model versions and performance.
• Serving endpoints to deploy models with ease, enabling real-time fraud detection via simple API calls.
• Databricks Apps for building a GenAI app in just one day—a process that would have taken weeks otherwise.
• Databricks Assistant for generating code and exploring data efficiently, enhancing productivity.

Using the open-source Meta Llama 3 and Anthropic Claude models, we extracted patterns from customer-agent chat transcripts to identify fraud signals, evaluate customer sentiment, and build adaptive models capable of learning from real-world scenarios. These insights not only streamline fraud detection but also provide our teams with actionable context behind each identified risk.

Vector databases further enhanced our GenAI capabilities by enabling precise searches across unstructured data. For example, when analyzing synthetic datasets generated by GenAI, we use these tools to uncover new patterns and refine our models faster than ever.

With Databricks’ tools, we can now build a GenAI application in only one day compared to seven days, a 7x reduction in development time. This boost in productivity and efficiency demonstrates that innovation doesn’t have to be time-intensive.

Forecasting the Future: From POC to Production

While this POC is still in its early stages, the potential impact is clear. The app has already shown its ability to identify fraud signals more effectively than traditional methods, offering real-time insights to agents and reducing fraud-related losses. The speed and adaptability of GenAI combined with the scalable infrastructure of Databricks will allow us to build on these early successes and integrate these capabilities into our broader fraud prevention strategy.

Our GenAI POC builds on our broader commitment to leveraging AI and data-driven solutions across our operations. As previously mentioned, our partnership with Databricks has already led to significant advancements in fraud prevention, reducing attacks by 80%. With GenAI, we are poised to achieve even greater results, including enhanced call center efficiency, real-time fraud alerts, and improved customer experiences.

What’s Next for GenAI at AT&T

Our first foray into Generative AI is just the beginning. This POC highlights GenAI's transformative potential in addressing complex challenges and delivering innovative solutions at scale. As we move forward, we plan to expand GenAI initiatives across customer service, agent training, and fraud prevention, ensuring that we stay ahead of evolving threats.

By building on the strong foundation provided by Databricks, we are confident in our ability to deliver exceptional value to our customers. Our next step is to scale this POC into a production-ready solution, integrating GenAI insights into our existing workflows to maximize their impact. The future of fraud prevention—and AI innovation—at AT&T is bright, and we’re excited to see where this journey takes us next.