Beyond Alert Fatigue: How Agentic AI is Reshaping Fraud Ops

If you have spent a day in the life of a bank fraud/AML investigator, you know that the biggest pain point comes when an alert lands and you start chasing its resolution. You spend 40 minutes across tabs and sheets chasing data only to realize that it is a false positive.

The industry is seeing turgid waters with 99% of AML alerts being false positives and only 2% of real-time fraud alerts qualifying as true positives. Investigators spend 90% of their investigative time collecting disparate data points as opposed to actually investigating. The problem cannot be solved by adding more rules or more predictive models. It can be solved by bringing in Agentic AI.

What is Agentic AI?

Contrary to popular perception, an Agentic AI does not simply report a risk score. Instead, it is a self-directed problem-solving agent that can formulate an objective based on a given prompt, collect relevant information from disparate sources, reason about the problem in context, and report its findings in a human-readable explainable format.

Let's take an example. If your salary account that usually sees ₹2 Lakhs per month in transactions receives and subsequently transfers out ₹20 Lakhs in an hour, traditional systems and even regular AI will raise a velocity/risk score. An agentic system goes beyond that and starts pulling in data from other sources. If it knows that ₹15 Lakhs came in from your parents and ₹5 Lakhs have been transferred to a commercial real-estate builder, it will provide you with an explanation that the money is being used to fund a home purchase.

How is Agentic AI being leveraged at Drona Pay?

Drona Pay has adopted an enterprise-grade multi-agent architecture to power its real-time transaction monitoring engine as well as case investigation workflows. Here's how the system is architected:

We have built a semantic layer/conductor that connects to legacy CBS and card switch systems, and we have deployed micro-agent workers that can handle various workflows in parallel.

The agents are purpose-built and help investigators in different capacities. For instance, while the Case Investigation agent helps in compiling context, the Interactive Rule Agent helps in writing rules in production-grade code. The Document Forgery Agent uses OCR to spot altered documents and claims. We have a web agent that helps in merchant onboarding by auto-profiling a website and predicting its MCC.

Critical to this is the privacy and grounding layer that makes sure that these agents work on isolated small language models to protect WPI data. We also have a RAG layer that helps in grounding responses in real-time regulatory handbooks and SAR guidelines.

Since these are autonomous systems, we have implemented a maker-checker approach wherein the first AI-driven output is always auditable by a human worker. These systems have hard-coded guardrails that respect the regulatory guidelines.

The impact of such a system can be seen in the numbers:

When implemented for a client, the system helped reduce the false positive noise by 85%, cut down case backlog by 80X, increase the investigation throughput by 4.5X, reduce the cost of operations by 70%, and enable 98% audit compliance.

Agentic AI is not about replacing investigators but augmenting them. By taking away the drudgery of data collection, fraud investigators can focus their time and effort on actually investigating.