Artificial intelligence (AI) may feel like the newest topic in banking, but financial institutions have used forms of AI for years. Machine learning, for example, already helps institutions analyze large volumes of financial transactions and identify activity that warrants closer scrutiny.
From machine learning to agentic AI
What is changing is the sophistication of the technology. Generative AI and agentic AI are creating new possibilities for financial crime professionals. Banks and credit unions considering AI in AML/CFT monitoring have an opportunity to make better use of data, focus investigators on higher-risk activity, and potentially create a more effective risk-based program.
AI does not remove a financial institution’s responsibility for its AML/CFT program. Human judgment, governance, testing, documentation, and risk-based oversight remain critical. But AI can help financial institutions deal more effectively with escalating AML/CFT compliance burdens, friction in workflows, and an ever-changing risk environment.
AML/CFT monitoring has always been a data challenge
AML investigators sift through enormous amounts of information. Transaction monitoring systems generate alerts based on scenarios and thresholds. Investigators gather customer information, review transaction histories and patterns, document findings, and determine whether further investigation or reporting of suspicious activity is warranted.
Financial institutions also continue to face staffing and resource challenges. According to a study by McKinsey & Company, approximately 20% of a bank’s full-time employees are dedicated to solving financial crime activities, with 60% relying on manual processes that are labor intensive. Insufficient resources can contribute to alert backlogs, inappropriate alert disposition, or monitoring that is not reasonably designed to capture the institution’s risks.
Technology can help manage this workload, but efficiency cannot be the only objective. Instead, financial institutions should also aim to use AI to improve their ability to identify and investigate potentially suspicious activity based on their unique risk profiles.
The Financial Crimes Enforcement Network (FinCEN) has encouraged responsible innovation in AML/CFT. The proposed AML/CFT program rule, published in 2026, discusses machine learning, generative AI, blockchain monitoring and analytics, digital identity, and Application Programming Interfaces (APIs) as innovative approaches that institutions may evaluate to combat financial crime more effectively.
Machine learning adds another dimension to monitoring
Traditional transaction monitoring relies heavily on rules, scenarios, and thresholds. Rules remain valuable because they are understandable, testable, and can be tied to known risks and typologies. But static rules may not identify unusual patterns outside established parameters.
Machine learning can complement rules-based monitoring by analyzing data for relationships, patterns, and anomalies that may be difficult to identify using predetermined thresholds alone. The Federal Reserve has described machine learning as a mature form of AI already used by financial institutions, including for fraud detection and prevention.
AI is not one technology, and using machine learning does not mean turning an AML program over to a black box and removing analyst judgment. Agentic AI has capabilities throughout the AML lifecycle, from identification, triage, investigation, and reporting, but human-in-the-loop must also be incorporated into these processes. Understanding the technology and being able to explain how it works to regulators is critical for the integrity of any AML program.
Machine learning can help identify unusual customer behavior, recognize patterns across large datasets, and prioritize alerts based on risk. The investigator still evaluates that information in context.
AI can help investigators focus on higher-risk activity
A typical transaction monitoring alert investigation may require an analyst to review customer information, previous alerts, account activity, counterparties, and transaction patterns. Gathering that information from multiple systems can constitute a substantial portion of an investigation.
AI can help AML investigators work more efficiently by identifying connections between customers and transactions, summarizing investigative histories, and prioritizing alerts. Machine learning can also support customer segmentation, automated scenario tuning, and alert scoring to help financial institutions focus resources on higher-risk activity.
Machine learning can also help address a longstanding monitoring challenge: false positives. A transaction may trigger an alert because it meets a scenario’s parameters even though the activity is reasonable for that customer. By considering historical behavior, customer risk, counterparties, and related activity, machine learning can provide additional context for prioritizing alerts, as well as addressing false positives.
Many financial institutions already use AI and machine learning for fraud detection, analyzing transaction patterns to identify potentially fraudulent activity. The same principle is relevant to AI in AML/CFT monitoring.
Reducing alert volume, however, should not be the primary measure of success. Institutions need to understand why alerts were reduced and whether monitoring continues to identify the risks it was designed to detect.
From generative AI to agentic AI
Generative AI and agentic AI are different technologies. Generative AI produces an output in response to a request. It might summarize a document, organize information, or draft text for human review. An example of this would be an AI-powered AML assistant that can auto-generate alert narratives that analysts can use as a starting point for further investigation.
Agentic AI goes a step further by taking action rather than simply providing information. An AI agent can determine the steps needed to complete a task, access relevant data, and carry out multiple activities with limited human intervention. For AML departments, this could mean automating portions of the investigative process while keeping investigators responsible for reviewing findings and making decisions.
Consider an alert involving unusual wire activity. An investigator traditionally might have to retrieve the customer’s profile, review previous alerts and transactions, research counterparties, compare activity with the customer’s expected behavior, and document the findings.
An agentic AI system could take the process further by coordinating multiple steps with limited human intervention. For example, it could collect and verify customer data, search external sources for adverse media, evaluate results based on risk and relevance, and prioritize cases requiring investigation. It could then generate case summaries, document its actions, and incorporate investigator feedback to improve future results. This would allow AML professionals to focus on higher-risk activity while maintaining oversight of the process.
Federal Reserve officials have discussed agentic systems capable of planning and executing multistep processes, although the technology remains in its early stages.
The investigator, or human-in-the-loop, remains critical. Agentic AI may change how information is gathered and analyzed, but human judgment should remain central to investigative decisions.
