Anti–Money Laundering AI Explained

Money laundering refers to the ways individual actors or criminal groups inject proceeds from their illegal activities into the global financial system to make them look like they were legitimately earned. US banks spend about $25 billion a year on processes to fight money laundering, and fines levied on banks worldwide for failing to prevent it topped $6 billion in 2023.

Criminals are getting more sophisticated in evading controls, while banks are finding it tough to pinpoint real money laundering actions because the great majority of alerts their tracking software raises for investigation are actually connected to benign transactions. Those false positives waste effort and money.

Now, financial institutions are starting to supplement or replace anti–money laundering (AML) software based on predefined rules with more sophisticated AI-powered software. This software is better at finding hidden patterns in transactions and relationships among people and companies, more thoroughly screens for suspicious activity, and more effectively scores customers based on their risk of money laundering. The result can be fewer false positive alerts, greater protection from illegal actors and regulatory fines, and lower compliance costs.

What Is AI?

Artificial intelligence is a set of statistical techniques that lets computers see relationships, make deductions, and predict scenarios based on patterns learned from large amounts of data. Financial services companies are using AI techniques to automate back-office processes, including fighting credit card fraud, personalizing product offers, making recommendations to sales teams, and combating money laundering.

What Is AML AI?

Traditional rules-based systems that look for warning signs of criminal activity or suspicious transactions based on preprogrammed patterns are yielding to AI-based systems that can detect the behavioral hallmarks of money laundering. Historically, AML software has looked for red flags that could indicate criminal activity, as well as complementary information such as a bank customer’s appearance on an international sanctions list, bank deposits just under the threshold requiring reports to the government, or transfers of amounts out of an account that are similar to those recently paid in.

The challenge is that criminals employ evolving tactics to launder their proceeds in what appear to be legitimate financial transactions. In addition to setting up shell companies to make ownership harder to trace, they invest in existing companies that do most of their business in cash and then inflate their revenues. They also deposit their cash in small amounts across several financial institutions and funnel cash through countries with lax regulations. That means that traditional AML methods are often ineffective while generating a very high number of false positives that can cost banks upwards of tens of millions of dollars a year.

AI-based systems can detect hidden transaction patterns among networks of people, compare behaviors with those that are historically common for an organization or its peers, assign risk scores to customers based on their past activity and Know Your Customer (KYC) information, and triage events to close investigations that are low risk. Fraud detection for transactions, electronic payments to vendors, AML, and KYC are among the top five AI use cases in financial services, according to research from AI chip maker NVIDIA.

Key Takeaways

• Artificial intelligence can help banks reduce their regulatory compliance costs by detecting more fine-grained changes in customers’ behavior and adapting to new risks as they arise.
• The software helps fight money laundering by finding previously hidden risks and reducing the number of false positive alerts AML teams need to investigate.
• Banks still maintain a similar number of reports for authorities on legitimately suspicious activity.
• The cost of inefficient processes is high, as global regulatory fines for failing to stop money laundering are rising.

Anti–Money Laundering AI Explained

Banks are under intense pressure to root out increasingly sophisticated money laundering techniques and avoid steep fines while keeping their regulatory compliance costs under control. By replacing rules-based software tools with AI-based AML applications, banks can improve their identification of suspicious activities by up to 40%, McKinsey & Company reports, while substantially reducing their number of false positives.

AI approaches include applying machine learning to scoring customers to predict their propensity to perpetrate a financial crime. AML applications also use unsupervised learning, in which a machine learning system isn’t shown labeled examples and gleans relationships itself from raw data, to identify changing customer behaviors and more accurately capture risk. AI systems can incorporate models of expected behavior, which flag deviations and replace fixed rules. AI-based AML tools also triage rules-based scenario events to automatically close or deprioritize low-risk investigations.

How Does AML AI Work?

When businesses or individuals want to open a bank account, banks perform risk assessments. That includes asking prospective customers a series of questions about their work, residence, income sources, and how they plan to move money. Banks also make sure prospective customers aren’t on international sanctions lists that would prohibit them from transferring funds. They also need to determine whether each individual is a so-called politically exposed person—a political figure or family member or close associate of one—which subjects them to more scrutiny. Banks then undertake KYC processes and score potential applicants for risk of money laundering or fraud.

The trouble is, some of the information depends on customers answering honestly, so financial institutions need automated ways of seeing if customers’ actual banking activity deviates from their stated intent. Traditional AML controls look at transaction data, including the international movement of funds, whether transactions of similar amounts are moving quickly among accounts, and whether customers have broken down large transactions into smaller ones. Criminals often move money to accounts in different countries whose AML regulations are less rigorous than where they reside. Another difficulty is that these behaviors can also have benign explanations.

AI-based systems are better at analyzing data to find patterns that human analysts and risk controllers alone can’t. The software can apply behavioral risk scoring to predict a customer’s propensity to commit a crime, run predictive models to show whether level-one investigations can be safely closed without escalation to more specialized teams, and simulate money laundering to assess the effectiveness of transaction monitoring systems. This can reduce the number of alerts that don’t signal actual money laundering activity, helping reduce compliance costs. Generative AI technology can help banks summarize initial assessments of risk and draft suspicious activity reports for law enforcement.

Commonly used AI techniques in AML include deep reinforcement learning, generative adversarial networks (GANs), and graph neural networks (GNNs). GANs generalize from examples of money laundering learned from training data to find modified patterns as criminals adjust their approaches. GNNs look for relationships among people and entities learned during training, including previously unidentified ones. That analysis helps banks spot money laundering activities involving groups of criminal actors. Deep reinforcement learning can teach AI models to learn about new relationships among data points by teaching the system to seek positive feedback for making the right decision. Models can then adapt their transaction monitoring to changing strategies.

AI-based AML systems can learn as they go. For example, if AI software finds transactions that share characteristics and are very unlikely to be money laundering, it can make recommendations for changes to the master system that would let similar transactions through going forward.

A 2022 Bank of England report on AI concluded that “one of the reasons AI is important is that it can enable new use cases”—for example, tackling the problem of synthetic identity fraud, in which criminals create identities “from a jigsaw of real data…which may be difficult to identify by human analysts.” AI-based AML systems can also employ unsupervised neural networks to look at very broad sources of data, including computer IP addresses and patterns of behavior, to generate alerts.

Benefits of AML AI

Traditional AML systems need to be adjusted for optimal sensitivity to activity that could raise red flags. Generating too few alerts risks missing criminal activity—and drawing attention and fines from regulators. Too many alerts can swamp banks’ compliance staff, who need to review each flag and decide how to act on it. AI systems have shown they can generate nearly the same number of suspicious activity reports (SARs) but with a greatly reduced number of false positives. Read on for more on these and other AI benefits.

• Increase/improve risk detection accuracy. AI techniques can help improve the accuracy of risk detection by consuming and synthesizing large amounts of structured and unstructured data, then learning about patterns of behavior and detecting anomalies. Neural networks can identify similar patterns to ones on which they’ve been trained, making recommendations that can close off criminal groups’ ability to make small changes to their fraud schemes to circumvent known rules.
• Reduce operating costs. AI-based AML systems can help reduce operating costs by lowering the number of false positive alerts that risk teams need to investigate. Each alert triggers a level-one investigation, which costs staff time. Between 90% and 95% of those alerts are closed before they’re escalated to a more intensive level-two investigation and SAR filing with authorities.
• Improve compliance and governance. Banks’ compliance departments, IT organizations, and business lines are under strain from changing AML regulations and a lack of global rules, though some convergence is starting to happen. Some regulators, including in the United States and United Kingdom, are encouraging banks to adopt AI for their AML systems. Financial institutions are also using AI and machine learning techniques to run risk-identification tests. AI can also help cut down on missed money laundering activity that can lead to regulatory scrutiny.

Limitations of AI in AML

Applying AI to fight money laundering won’t work well if banks don’t have enough high-quality data to train models to deliver consistently accurate results. Banks also need to make sure they have the right talent on staff to train, fine-tune, and maintain AI models, and they need to consider customers’ data privacy when designing systems. AI can also be opaque: It’s not always clear how a generative AI system arrives at its answers. McKinsey advises banks to meet with regulators well in advance of developing an AI-based AML system. Read on for more on these and other AI limitations.

• Data Quality and Availability. Incomplete or inaccurate data can hamper the performance of AI models, which need to see enough high-quality examples during their training to be able to accurately flag suspicious transactions once they’re deployed. Banks with limited access to data, especially data on true examples of money laundering, may wish to consider another approach.
  
• Regulatory and Compliance Challenges. Ever-changing and sometimes vague and inconsistent AML regulations can tax banks’ compliance departments, IT organizations, and lines of business. For instance, recommendations from the Financial Action Task Force (FATF), a global standards-setting body for AML, leaves wide latitude in its guidance to oversight authorities. JPMorgan Chase CEO Jamie Dimon has called for AML requirements to be simplified and improved. The rules are starting to be unified. In 2024, the EU created a new AML authority based in Frankfurt whose rules apply to firms throughout the bloc, without the need for them to be transposed into national laws.
  
• Operational and Technical Issues. Most banks also need to tackle systems integration challenges, since data stored in mainframes and other legacy systems isn’t usually ready for processing by AI. Banks also struggle with keeping sufficiently high-quality data on their customers, particularly longtime ones whose history may be partially stored in nonstandard formats and taken from pen-and-paper forms, according to consultancy McKinsey & Company.
  
• False Positives and Negatives. Rates for false positives—benign financial transactions that get flagged by software as potential money laundering activities—can run as high as 95%. But banks still need to look into them, which is a costly and time-consuming process. Conversely, false negatives—when transactions by sanctioned entities or real money laundering slip through—can lead to regulatory actions and reputational damage.
  
• Adaptability and Evolution of Criminal Tactics. Money launderers are sophisticated adversaries who employ ever-evolving techniques to evade detection. Once bad actors identify rules used to detect laundering, they can slightly modify their behavior to evade them.
  
• Privacy Concerns. Banks designing AI to combat money laundering need to consider the privacy aspects of these systems. Banks have access to significant amounts of data, including personal information, and the rights to use this information must be evaluated in context.

AML AI Use Cases

Banks are applying AML AI techniques when bringing customers onboard, monitoring their banking activities, and reporting suspicious behavior to authorities. The software can make processes faster and more effective by building behavioral profiles from sometimes-hidden patterns to better classify transactions, scouring documents and news for potentially risky customers, and speeding regulatory report writing.

• Transaction Monitoring. AI models can monitor transactions for illegal activity in two main ways.
  • Pattern recognition. From their training data, AI models can learn to recognize transaction patterns that evade traditional rules-based AML systems—for example, identifying “structured” transactions, in which large sums of money are broken up into smaller amounts, or analyzing large amounts of data to identify shell companies used to transfer money. AI can also model expected customer behaviors, as well as deviations from those behaviors that can indicate criminal activity, replacing rules-based approaches that may not be as accurate.
  • Real-time monitoring. The speed of digital payments is spurring demand for AI-powered AML systems that can sift through large volumes of data quickly, even in real time. Michael Hsu, acting US comptroller of the currency, said in a January 2024 speech that faster digital payments are leading to faster fraud, pressuring banks to build “the right brakes for a more real-time financial system.”
• Customer Due Diligence (CDD) and Know Your Customer (KYC). Banks can identify and screen customers online using AI-based automated onboarding techniques to help make KYC processes faster and more accurate. These include digital identity verification and scanning ID documents. Through the continuous monitoring of transactions, banks may better identify high-risk customers by analyzing more data than periodic reviews allow.
  • Automated onboarding. Banks can help improve speed and accuracy when opening customer accounts by scanning identification documents for online verification and applying AI to assess authenticity.
  • Continuous monitoring. Because customers can behave differently over the course of their relationship with a bank, and because election outcomes globally can change whether someone is considered a politically exposed target, financial institutions are employing AI-based tools to continually check transactions, beneficial ownership, sanctions lists, and media coverage. Continuous monitoring checks if customer behavior may have become riskier since a bank last assessed it.
• Suspicious Activity Reporting (SAR). Banks need to file SARs to regulators to flag suspected cases of money laundering and terrorist financing.
  • Automated reporting tools enabled by generative AI can help generate SARs more efficiently than human analysts working alone.
  • Enhanced reporting accuracy. Many SARs suffer from unclear narratives and missing information, leaving ample room for improvement using generative AI systems, which can also create lists of follow-up items.
• Sanctions Screening. Frequent updates to international sanctions lists (such as after the beginning of the Russia-Ukraine war) and difficulties matching business entities across variations of people’s or companies’ names in different countries and languages can overwhelm traditional AML systems.
  • Automated screening using AI can extract and classify information from unstructured documents, find synonyms for red-flag terms, and discard similarly spelled terms with different meanings.
  • Reduction of false positives. The result can be a reduction of false positives, requiring fewer time-intensive reviews by analysts that add to costs.
• Enhanced Analytics and Visualization. Data visualization techniques, including graphs of relationships among people and entities, can help nontechnical business users see changes in risk and the geographic distribution of suspected money laundering cases.
  • Data visualization. In addition to relationship graphs, AI techniques can help analysts spot the location of nefarious activities on maps and drill into dashboards to access finer details. The result can be quicker, more effective decision-making.
  • Dashboard reporting shows metrics and progress against key performance indicators (KPIs) for transactions monitored, alerts generated, SARs filed, and investigations opened and closed.
• Regulatory Compliance. AI tools can help financial institutions stay abreast of and adapt to regulatory updates, including with documentation that enables auditability.
  • Regulatory updates. Rising money laundering and KYC fines, and the new regulatory bodies meant to enforce them, are increasing the demand for technology that can help banks stay on top of changing rules and reduce the likelihood of fines by replacing or complementing manual processes with automated ones.
  • Audit trails. These software tools can also generate audit trails that demonstrate how AML decisions were made and help write logs for audits that show activity and data access.
• Case Management. Workflow automation and collaboration tools can help with AML compliance by prioritizing alerts, recommending actions, and automating reporting. Software can track alerts and provide dashboards displaying suspicious activities.
  • Workflow automation. Financial institutions may be able to save on average a quarter of their yearly compliance costs using workflow automation tools against financial crime, according to IT consultancy KPMG. AI tools can also keep KYC processes current by bringing regulatory changes into onboarding workflows.
  • Collaboration tools. Case management software helps departments coordinate their activities by providing a central storehouse of compliance information.
• Fraud Detection and Prevention. Banks have some time to stop money laundering after it’s discovered, though fraud cases should ideally be stopped before transactions go through to avoid losses. AI systems employ adaptive learning, so they can can help with transactional screening. AI models also afford a broad view of customers, so banks can see investigations and reports that span AML, fraud, and sanctions for bribery and corruption, according to KPMG.
  • Integrated fraud and AML solutions. Combining techniques to counter fraud and money laundering in a single software package benefits banks by helping anti-fraud and AML teams collect and share pertinent customer data, present an overview of risks facing the bank, and close off loopholes criminals may use to evade notice.
  • Adaptive learning. As banks confirm new cases of fraud, AI systems can use this data to improve over time, particularly when it comes to finding edge cases near the threshold of detection that are likely to result in missed activity or false positives.

How to Incorporate AI into AML

Banks that want to redesign AML processes to incorporate AI techniques should first assess their data strategy, including the data they have. They need to consider how AI can be used across departments and workflows handling KYC, customer onboarding, and anti–money laundering. The resulting systems need to be assessed in context for suitability and evaluated for regulatory compliance. Read on to learn more about these and other steps.

1. Assess current AML processes. Banks need to review how they counter money laundering, how effective current systems are at preventing it, and the costs that could be saved and the improvements that could be made from an AI-based approach.
2. Define objectives and requirements. Banks should scope out the goals of an AI implementation, including with clearly defined success criteria, such as lowering operating costs and reducing the number of false positives.
3. Shore up data collection and preparation. Banks need to ensure their data is clean and of a high enough quality and quantity to train AI models. They also need to have sufficient data science talent on staff to tune models and refine approaches (more on that later).
4. Choose the right AI tools and technologies. Selecting an AI system that fits the required use cases is key. Systems should be capable of monitoring transactions in real time, applying machine learning and natural language processing (NLP) to customer onboarding and KYC processes, and using GenAI and NLP to help generate suspicious activity reports. Banks can use predictive analytics to evaluate anomalous or suspicious behavior. Graph AI analysis can help find networks of people and entities that aren’t apparent to analysts.
5. Develop and train AI models. There are two main ways AML AI models can be trained. With supervised learning—for behavioral models, customer risk scoring, and event scoring for AML and sanctions lists—models are shown labeled examples from which to learn. This is advantageous when a model needs to learn about relationships between inputs and outputs. For use cases such as customer segmentation and anomaly detection, banks commonly use unsupervised learning, in which models are shown unlabeled cases of money laundering and fraud, as well as false positives. Without any help from data scientists, the model learns to identify the characteristics of both groups of transactions. Unsupervised models can learn about relationships in data that haven’t been spotted before.
6. Integrate AI with existing systems. Many AML processes run on legacy IT systems, so banks need to invest in building connectors with older transaction monitoring and reporting systems and their various data types, or modernizing their infrastructure to handle the demands of AI.
7. Train and support staff. Training on AI tools and processes is important both for regulatory compliance and for overcoming staff resistance to adopting the technology.
8. Strive for continuous improvement and adaptation. AI is designed to promote continuous learning, and banks need to embrace a similar mindset while deploying and using the technology. For AML alerts, banks also need to consider the so-called recall ability of models, which measures their capacity to generate nearly the same number of actual suspicious activity reports from a much smaller group of alerts.
9. Help ensure regulatory compliance. Laws and regulations regarding AI are evolving, and banks need to stay on top of them to maintain compliance. Internal controls, training, and appointing an officer for managing ongoing compliance are all key parts of effective AML compliance.

Future of AI in AML

AML requirements set by global authorities are evolving, while banks’ AML budgets are under pressure, making AI-powered analysis and automation more attractive. In the US, the Treasury Department's Financial Crimes Enforcement Network (FinCEN) is considering rules that would extend the Bank Secrecy Act to investment advisors, including requiring them to file SARs. Swiss regulator FINMA has been ordering banks to conduct more thorough AML reviews, and the EU’s new Anti–Money Laundering Authority is expected to introduce more direct supervision of up to 40 financial institutions.

Banks may look to double down on new AI-enhanced methods of fraud detection, as hiding money behind high-end cars, collectibles, jewelry, and art has made the practice harder to trace. Money launderers also use social media to recruit low-ranking workers to deposit money, making nefarious activity harder for traditional systems to root out.

Modernize and Strengthen Your Anti–Money Laundering (AML) with Oracle

Oracle Financial Crime and Compliance Management Cloud Service includes software engines that dynamically score entities and transactions for risk and freeze transactions to sanctioned entities or countries for immediate review by an analyst. It also has a full-featured case management capability designed for the way investigators work.

Oracle Financial Services Compliance Studio includes statistical analysis and supervised and unsupervised learning AI technology, which helps better understand and monitor risks and reduce the costs of maintaining a platform for compliance and countering financial crime.

Oracle’s approach helped one large multinational bank implement AI models within six weeks and generate 45% to 65% fewer alerts while still producing at least 99% of the number of suspicious activity reports it had when its alert volume was much higher.

Oracle Financial Services Compliance Agent is an AI-powered cloud service that lets banks test their transaction monitoring systems and simulate bad actors to stress test their AML programs, helping to reduce costs and regulatory risk. Oracle is also developing a generative AI component for its financial crime software to help write case narratives for reports.

AML AI FAQs

Will AML be automated?
Banks are increasingly automating their AML processes using AI tools that can gather and process data from across departments. These tools complement and support the analysts and other workers who are responsible for the AML function.

What is generative AI in anti–money laundering?
Banks are using GenAI to look for related terms that aren’t hard-coded into the rules engines of traditional AML software, to identify tough-to-spot relationships among transactions, and to create narratives for suspicious activity and other reports.

What is intelligent automation in AML?
Intelligent automation is used to reduce the manual work involved in reviewing transactions an AML system incorrectly flags as fraudulent. It does that by applying new patterns the AI model learned to classify future transactions. This can help lower banks’ costs and improve accuracy.