Agentic AI for AML: The Investigative Process Before and After AI Agents

Your AML team became analysts to solve financial crimes, but they're burned out from alerts that shouldn't have reached their desk in the first place. It's no surprise that analyst turnover in AML is notoriously high, with most leaving within their first year.  

The real impact isn't just lost time; it's compromised investigation quality. When analysts suffer fatigue from false positives, they lack the focus to analyze the cases where complex financial crimes are taking place. But there's a better way forward.

What is agentic AI and how does it work? 

Traditional AI analyzes structured data and makes predictions. Generative AI (GenAI), can create new content from a prompt. As highlighted in our recent Hawk Guide to Agentic AI for Financial Crime and Compliance Leaders, Agentic AI is different. It’s not just one monolithic AI model, but an orchestrated set of tools and models working together, each of which is often referred to as an “agent". For example:
 

• A Data Gathering Agent collects information from disparate sources
• A Typology Agent studies the data for risks and labels the case accordingly
• A Narrative Agent writes the SAR for the analyst to review for accuracy  

When Analysts struggle with "alert fatigue", they stop looking thoroughly at alerts due to all the false positives. This creates a big operational risk where real financial crime is allowed to slip through. AML agentic AI models significantly improve alert quality and reduce false positives by automating the initial triage.  

Here’s how they do it:
 

• Access multiple data sources to create a comprehensive case assessment
• Help analysts quickly grasp the core risk without sifting through raw data, enabling faster resolution of false positives and quicker escalation of true risks
• Recommend optimal alert queue ordering based on analysis of past resolution times and outcomes. By looking at historical data to see which alert types and patterns are most likely to be true positives, the agent can recommend moving high-priority cases to the top of the analyst's queue 

In short, agentic AI makes recommendations based on real-time data with minimal human input. But it still allows a human-in-the-loop approach that complex and high-risk cases require. In this article, we’ll compare how the end-to-end investigative process looks before and after introducing agentic AI. 

How AI agents streamline the investigative process  

Automates data collection  

An AML investigation begins when an alert triggers on an account. The analyst manually opens the case and scrutinizes whether the red flag needs pursuing. If the alert appears valid, they create a case and assign it for investigation.  

This is where the real inefficiency begins. Analysts must collect customer details from wherever they exist. Some information already sits in customer and transaction databases, but additional data requires analysts to search outside systems: AML databases, credit bureaus, and even Google searches. Each source requires separate login credentials, different search methods, and manual documentation. 

A data-gathering AI agent, on the other hand, automates the information gathering process. It rapidly pulls data from transaction databases, customer records, alerts in surveillance systems and historical investigations to create a complete, 360-degree view of a case. Automating the research process with agentic AI improves both efficiency and accuracy, enabling more thorough and timely case assessments. 

The agent accesses these data sources through a few methods: API integration with dedicated adapters to pull data from an FIs case management system, UI scraping to read data from the case management interface, or Modular Connectivity Protocol (MCP) servers that customers can provide to grant secure access to data sources.  

The agent then creates a concise, structured case overview with key risk indicators and red flags upfront, so analysts can quickly dismiss obvious false positives. 

Identifies and labels cases by typology  

Next comes the detailed transaction breakdown. Analysts must sort through every transaction on the account, searching for anomalies, suspicious money movements, and unusual behavioral patterns. But the sheer amount of data makes this time-consuming. They spend hours manually connecting disparate data points across multiple sources, relying entirely on their experience and intuition to identify suspicious patterns. They must then figure out why the behavior triggered suspicion and determine what type of financial crime they're investigating. 

The AI agent analyzes collected data against thousands of known financial crime typologies, automating this entire process. Instead of spending hours on pattern recognition and crime classification, analysts receive cases that are already labeled with specific typologies, giving them immediate clarity on the alert's nature and allowing them to focus on investigation rather than identification. 

Provides investigative guidance  

The AI agent also acts as an intelligent co-pilot, suggesting the next best steps and eliminating the guesswork that slows investigations. Rather than analysts deliberating which systems to check or what queries to run, the agent recommends specific actions based on the identified typology. This ensures every case follows a consistent, compliant approach while dramatically reducing investigation time. 

Captures decision rationale  

Comprehensive documentation is critical: analysts must detail their findings and actions to keep other departments informed while making future investigations easy if alerts trigger on the same accounts. But different analysts interpret and document findings differently, leading to inconsistencies that can’t be explained to auditors and regulators. 

The AI agent standardizes rationale capture for every case, making it audit-ready. This eliminates the need for analysts to manually write detailed summaries, as the agent automatically records the reasoning behind each decision. This ensures all case files meet internal and regulatory requirements, reducing the risk of audit findings.  

Drafts SAR narratives  

When an investigation confirms financial crime has taken place, the analyst must file a SAR. This final step often becomes another time sink. Analysts write SAR narratives from scratch, spending hours crafting reports that meet regulatory requirements. Back-and-forth reviews for narrative approval stretch the process even longer. 

Agentic AI streamlines this workflow by drafting initial SAR narratives that meet regulatory standards. It reduces revision cycles, improves report quality, and guarantees faster alert-to-SAR completion times. Instead of staring at a blank document, analysts receive a first draft to review for accuracy. This automation shifts their focus from writing to refining, using their expertise to augment the narrative. 

Hawk: An AML Agent made for you 

Every financial institution has its own unique workflows and policies.  

Hawk's AML Analyst Agent is designed for maximum flexibility, with an orchestrated agentic workflow that is fully customizable. Its Agentic Deployment Management platform has an easy, drag-and-drop interface. This lets compliance teams upload their SOPs and policy documents, then design and change investigation workflows based on those documents without writing any code.  

It can also connect with external components and models (MCPs). But more importantly, AI agents can be configured to pause at specific points in the investigation process to get human input, review, or approval. This granular level of control lets you choose the best model or require human judgment at certain steps in the investigation, making the process both faster and more accurate. 

Overlay by design 

Hawk’s AML Analyst Agent works as an overlay, so it can run on top of your existing AML infrastructure without having to replace any current systems. It uses API-based integration to smoothly access data sources, but if a full API setup isn’t practical, it can use UI-based integration. This means the agent can work directly with your existing screens, reading and entering data like a human analyst would.  

Auditability & explainability in its DNA 

The Hawk agentic solution is built to give a full, clear record of its investigative process, which is crucial for meeting regulatory requirements. It keeps a detailed log of every action, including timestamps and model calls, complemented by a confidence score showing how certain the agent is. Most importantly, it creates a visual map of its decision-making process, so investigators and auditors can easily follow the logic back to the original data, with citations for every source the agent used. 

Download our agentic AI whitepaper  

What you've read here is how agentic AI transforms the end-to-end investigative process. But it can do even more to improve the accuracy of your overall AML system. Implementing agentic AI requires careful planning, buy-in from key stakeholders, and an understanding of important factors when choosing an agentic AI solution. Our latest whitepaper provides a full roadmap, including use cases, evaluation criteria, operational fit, expected ROI, and a practical buyer’s checklist. 

Download your copy below.