In movies, investigators put the pieces together. With a massive wall of mugshots, maps and post-it notes, they connect the dots using years of honed intuition—and red yarn. It is an analytical process, and it is exciting.
In reality, investigators, especially anti-money laundering (AML) investigators, are spending much more time finding the pieces than figuring out how they fit together—nine times as much time, in fact. They are not analyzing evidence; they are shuffling through 10 browser tabs, from LexisNexis to Google searches, compiling hundreds of screenshots into a comprehensive, by-the-book trail of compliance documentation for each case. They do not have much time for drawing conclusions when they have to spend hours copying information field by field into Microsoft Word and Excel work papers.
In reality, investigators are not doing much investigating.
The Real Enemy
Every day, AML investigators take on some of the most difficult challenges faced by anyone in the financial services industry. Their job includes daunting tasks like unraveling the rapidly evolving techniques used by criminals and coordinating closely with multiple law enforcement (LE) groups to build cases against bad actors. Yet these are far from the hardest aspects of their role.
Instead, the most difficult part of an investigator’s job is dealing with the shambles of investigations technology today, which is leading to systemic failure. The United Nations Office on Drugs and Crime has estimated that the cumulative global AML effort catches less than 1% of the more than $2 trillion laundered annually. In other words, the current process is producing a 99% failure rate.1
The process is broken and it is failing to achieve the results society needs.
More Menial Than Meaningful
To understand the problem, one needs to take a look at the investigations process itself.
The typical investigation involves pulling information from 10 or more sources, manually copying the data into work papers and fixing granular formatting errors. It is only after hours of this that the actual review begins. In short, a sea of paperwork stands between every investigator and their actual job. In the words of a senior AML executive, “I wish I could spend my day fighting financial crime, but instead I spend it filling out spreadsheets.”
This is not an effective use of talented minds in the space; it is endlessly repetitive, duplicative and demoralizing. It is not work fit for people—but it is perfect for computers.
The Data Problem
While the time wasted in the current process is significant, it is not the most impactful aspect on AML program effectiveness. Rather, data lies at the core of the 99% failure rate issue. The inability to do anything with the data investigators have on hand is the true limiting factor.
Put plainly, the current process is a bad data factory, making any attempts at applying data science to the effort near impossible. Clean, organized and well-labeled training data is key to machine learning. Unfortunately, in the AML world, any data that could be used to build models are fragmented across documents, spreadsheets and SharePoint folders and share no common standards.
Even if an AML department is lucky enough to have data scientists available to help, this effort would soon be stalled by an imposing wall of spreadsheets and data cleansing tasks long before any meaningful progress can be made.
Robotics Is Not the Answer
In the words of Bill Gates, “The first rule of any technology used in a business is that automation applied to an efficient operation will magnify the efficiency. The second is that automation applied to an inefficient operation will magnify the inefficiency.”
In the context of the AML world, this means that speeding up a process that produces a 1% success rate will only speed up failure. Process automation is an essential tool to deploy, but there are bigger steps that need to be taken.
Solving the 99% failure rate issue requires a comprehensive rethinking of how AML is approached. While efficiency is certainly a problem plaguing AML investigations, there is a more fundamental issue at play here: information asymmetry.
In essence, criminals are agile, continually evolving and freely share techniques amongst themselves. However, the processes the good guys use take much, much longer to develop. It can take months or years for information about new typologies to spread across the industry. In short, the technical advances happening on each side are a different order of mathematical function. In the battle between linear and exponential change, this is no exception to the rule.
The advent of “crime-as-a-service” serves as an excellent example of the challenge AML professionals face. Over the past five years, some sophisticated criminals realized they could earn more at a lower risk if they did not move the funds themselves. Instead, they test industry controls, build tools and write guides on how to exploit vulnerabilities that are then sold to less technically-savvy criminals on the dark web.
This is a risk shared by most financial institutions. If an AML team at any bank searched the dark web for their institution’s name with a burner computer, they would likely find multiple highly detailed (and probably high-quality) resource guides detailing how to defraud or launder money through their organization for sale.
In our experience, inexplicably strange behavior in an account that could never be directly tied to fraud losses or laundering behavior would occasionally be traced to bad actors testing the control environment for vulnerabilities that they could then sell. This gives a completely new meaning to “transactions with no apparent economic, business or lawful purpose.”
In a world where criminals can freely use bleeding-edge technology to attack while AML departments are defending with spreadsheets, COBOL databases and rule-based monitoring systems, it is abundantly clear why the good guys are not winning the fight.
If the AML effort is going to catch up, it needs to focus on getting relevant information to investigators as quickly as possible. Process automation is a part of the solution, but the first step is to ask what those processes would look like if they were redesigned from scratch using modern technology. If AML is going to be agile enough to put the criminal world back on its heels, it needs a transformation.
The Future of Investigations
Any speech on the introduction of artificial intelligence (AI) in AML given at a recent conference invariably leads to questions from the audience on the impact it will have on jobs.
The response should be, “AI is not here for our jobs, it is here to let us focus our time on what the human brain does best, make intuitive connections when presented with the right information.” There will always need to be a human in the loop. AI is there to empower the investigator.
In a world where criminals can freely use bleeding-edge technology to attack while AML departments are defending with spreadsheets… it is abundantly clear why the good guys are not winning the fight
So what does “empowering investigators” actually mean?
The following examples illustrate potential applications of AI that are under development across the regtech world:
- Alert intelligence: Systems analyze alerts to see if patterns or components of cases have been observed before and present relevant information to the investigator. No more reliance on the investigator’s memory or clunky database queries to find patterns; the system flags potential ties and lets the investigator decide if they are significant. Importantly, the system learns from actions taken across the entire team, and potentially across the whole industry, allowing everyone to benefit from the expertise of others to understand what they are seeing.
- Contextual information: Once an alert is found to be valid, the system prepopulates the case file with relevant information from the institution’s internal databases and any third-party sources that it can access. Consequently, analyst time that was once spent on gathering data today is freed up to focus on the actual investigation.
- Natural language processing: To date, natural language processing (NLP) has focused mainly on saving time by generating investigation narratives. In the future, it will extract relevant information from text-dense documents, customer chat, email and call transcripts, social media communication, dark-web posts and even media sources, helping investigators separate the signal from the noise. Further, in a process known as information extraction and linking, NLP models can be trained to recognize key intelligence (e.g., names, dates, events, facts) to add to the system’s ability to surface relevant contextual information.
- Secure collaboration and information sharing: The development of open and interoperable data standards will enable working together across companies and international borders. Privacy-enhancing technologies (already under testing by international regulators) will ensure that personally identifiable information and other confidential information is not compromised while permitting behavioral modeling to become a reality.2 Anonymized suspicious behavior models are sharable, enabling learning new typologies the moment any investigator finds one.
- Feedback loops: Investigation results are no longer relegated to a SharePoint black hole; they are now directly fed back into transaction monitoring models to improve the accuracy of alerts in real time. LE can directly inform monitoring systems across the industry to flag priority typologies.
- Measurable effectiveness simplifies the regulatory process: Exam teams are now able to see how well an AML function works by measuring its effectiveness (including LE feedback) rather than its processes.
What Could This Mean?
For the investigator, the advent of AI-support will be truly revolutionary; one only needs to look to the world’s most sophisticated fintechs to have a grounded vision of the future. The best of them have been deploying machine learning into their investigations process for several years now. Some possible results include the following:
- There is an inversion from the industry average of greater than 90% false positives to less than 10%.
- Teams have the time and tools needed to do the deep investigative work required to unravel the most sophisticated patterns.
- The delta between evolving criminal typologies and the ability to catch them is significantly reduced.
Back to First Principles
Perhaps the best way to think about the future of investigations is to apply first principles thinking.
Elon Musk, a famous advocate for this approach, has put this method at the center of how he thinks about all of his businesses. He has said on the topic, “Focus on signal over noise. Do not waste time on stuff that does not actually make things better.”
The development of open and interoperable data standards will enable working together across companies and international borders
When creating a vision of the future, it is important to evaluate whether the current approach is actually achieving the desired results from an objective point of view. If not, exploring new approaches is fundamental. It will not be easy to obtain the (admittedly) starry-eyed vision for the future outlined above. Still, there is good reason to believe that the application of AI will be the key to taking a meaningful chunk out of the remaining 99% of uncaught financial crime.
Matt Van Buskirk, co-CEO, Hummingbird Regtech, Washington, D.C., USA, email@example.com
Steve Cohen, COO, Basis Technology, Cambridge, MA, USA, firstname.lastname@example.org
- Preeta Bannerjee, “UNODC estimates that criminals may have laundered US$ 1.6 trillion in 2009,” United Nations, October 25, 2011, https://www.unodc.org/unodc/en/press/releases/2011/October/unodc-estimates-that-criminals-may-have-laundered-usdollar-1.6-trillion-in-2009.html
- “2019 Global AML and Financial Crime TechSprint,” Financial Conduct Authority, July 29-August 2, 2020, October 25, 2011, https://www.fca.org.uk/events/techsprints/2019-global-aml-and-financial-crime-techsprint