The telecommunications industry has always grappled with its biggest nemesis – fraud. From SIM box fraud and Wangiri scams (explained below) to device and subscription frauds, the sector loses billions annually to increasingly sophisticated schemes.
However, telecom companies are putting up a good fight with AI for fraud detection and prevention.
Virgin Media O2’s AI-generated grandmother, Daisy, is an AI system that engages with scammers through natural conversations. It keeps them occupied for up to 40 minutes while telling stories about grandchildren and displaying calculated tech illiteracy.
The system has already conducted over 1,000 conversations with potential fraudsters, effectively turning the tables on scammers by using their own tactics against them.
Fraud not only causes financial losses but also erodes customer trust. “Fraud has evolved—it’s now more sophisticated, and there’s more data to handle. Rule-based systems can no longer keep up. This is where AI comes in,” Harsha Angeri, VP & head of AI business at Subex, told AIM.
Telecom fraud is a global issue, with annual losses estimated at $32.7 billion. Fraudsters are leveraging advanced technologies like VPNs, SIM boxes, and even AI itself to bypass traditional security measures. A Reddit user recently remarked, “…financial fraud and cybercrime are off the charts. I don’t know anybody in my office or family who hasn’t received a scam call in the last few months.”
How AI is Helping Telecoms Detect Frauds
One of the most prevalent types of telecom fraud is SIM box fraud, where international calls are rerouted as local ones to evade tariffs. Previously, it was easy to flag suspicious SIM cards by identifying patterns like only outgoing calls and no incoming calls. However, as Angeri points out, “Today, fraudsters use mobile SIM boxes that move around in vans to avoid detection. It’s no longer as simple as writing an ‘if-else’ rule.”
AI models now analyse billions of call detail records (CDRs) daily to detect such sophisticated patterns. This capability is crucial in countries like India and Indonesia, where telecom operators process hundreds of millions of calls every day.
Another major issue is device fraud, particularly in regions like the Middle East and the US. In these markets, customers often purchase high-end devices on instalment plans. Fraudsters exploit this system by paying one instalment for an expensive device like an iPhone 16 Pro before disappearing.
Angeri explained how AI tackles this: “We assess risk by looking at features like customer tenure, payment history, location consistency, and even whether the customer recently switched SIM cards.” These insights enable telecom operators to flag high-risk transactions before they occur.
Wangiri scams are another significant challenge. Here, victims receive missed calls from premium numbers and are charged exorbitant fees when they return the call. Differentiating between fraudulent missed calls and legitimate ones (such as flash calls used for OTP verification) requires advanced anomaly detection techniques.
“AI helps us differentiate these cases by examining call metadata and usage patterns,” Angeri said.
For example, Subex uses supervised learning models like Random Forests and XGBoost to analyse hundreds of features—such as call duration, frequency, and location—to flag suspicious activity.
AI also addresses newer challenges like flash calling, where social media platforms use missed calls for user verification without paying traditional SMS charges. While not strictly fraudulent, it represents revenue leakage for operators. AI models help identify such behaviour effectively.
The scale at which these models operate is staggering. In Indonesia alone, some operators process up to 150 billion call records daily. As Angeri emphasises: “You can’t do a hobby project here. The scale is enormous. AI models must handle billions of records in real-time.”
Why Legacy Systems are Still Relevant
Despite AI’s advancements, legacy rule-based systems continue to play a crucial role in telecom fraud management. These systems provide foundational insights that complement AI-driven solutions.
Subex’s legacy system, ROC (Revenue Operations Center), is still widely used alongside its new-generation AI platform, HyperSense. “Legacy models provide seed data and initial insights that help train AI models effectively,” Angeri said.
For simpler fraud scenarios—such as identifying SIM cards with unusual call patterns—rule-based systems remain cost-effective and reliable. Legacy systems also ensure continuity during the transition to AI-powered solutions.
Many telecom operators prefer a hybrid approach that combines the stability of traditional systems with the adaptability of AI models. This approach allows operators to gradually scale their AI capabilities while leveraging existing infrastructure.
Angeri highlighted another key advantage of legacy systems, which is their ability to handle privacy-sensitive data securely within telecom operators’ infrastructure. “Our systems are deployed within the telco’s environment,” he said. “Data like CDRs or billing records never leave their secure systems.”
This ensures compliance with stringent privacy regulations while enabling robust fraud detection.
However, legacy systems alone cannot tackle today’s sophisticated fraud schemes. “Fraud detection is an ongoing process requiring frequent model retraining as patterns evolve,” Angeri said. Integrating legacy systems with advanced AI techniques ensures telecom operators can effectively address both traditional and emerging threats.
“Fraud will happen—that’s the reality. But as soon as patterns emerge, we must catch them so they don’t propagate,” he concluded. With $40 billion lost annually to telecom fraud globally, adopting such hybrid solutions isn’t just an option—it’s necessary to safeguard revenues and maintain customer trust in an increasingly digital world.
The post AI vs Fraud: How Telecom Giants Are Outsmarting Scammers appeared first on Analytics India Magazine.
