The end-game of anti-money laundering operations will, logically, be a fully automated process whereby monitoring and screening, alert investigation and decisioning, and generation and filing of regulatory reports are executed without human intervention. This should be feasible within ten years, and perhaps sooner.
From our current viewpoint from within a paradigm that relies heavily on manual operations, automated AML may seem like a distant goal. But the reality is that we are already halfway there. To state the obvious, transaction monitoring and screening — the two main building blocks of AML — are performed by automated software engines. Although model building and tuning require significant and ongoing effort, once a scenario or filter is put into production the engines hum along, perform their analysis, and populate case management systems with the resulting alerts, risk scores, and associated transaction and customer data.
What is not automated today is the investigation of alerts once they've entered the case management system. And therein lies the rub, because the majority of alerts are false positives. False positive rates may be 70% or higher, and this immediately gums up the operation by draining resources from the crucial and value-added work of identifying true positives. High FP rates are the main driver behind the massive expansion of AML back- office teams, which have grown to be as large as small cities.
Generating time, cost savings
The next step in AML automation, then, is getting machines to handle the disposition of the obvious false positives that make up the majority of alerts. The good news here is that progress is being made in applying technology to this problem. Rules-based machine learning can suppress alerts consisting of data that has been seen and passed before. Going a step further, artificial intelligence can analyze previous alert decisions to understand the characteristics of false positives and predictively apply these models to the decisioning of new alerts. Anecdotal evidence suggests that techniques such as these can generate time and cost savings of 30% to 70% by automatically separating the wheat from the chaff of false positives.
That leaves the remaining alerts that must be investigated by human analysts. Here, technology is enabling a cyborg operational model where analysts and machines work interactively. A good example is entity investigation, which is integral to KYC workflows and also a common element of transaction monitoring alert investigations. Here, robotic software can do much of the heavy lifting by searching external data and media sources and compiling the relevant found information into ready-to-go entity profiles to be presented to analysts for review. Additionally, artificial intelligence can be used to not just search for keywords but to understand contextual information about the entity and generate more meaningful entity profiles.
In the case management context, visualization tools have long been important in providing graphical analyses of data for analyst review. Visualization tools have continued to evolve, to the point where they now provide a spectrum of value added information and insights, much of which would elude an analyst relying on table-driven compilation of data. For example, link analysis tools are now backed by the power of graph analysis, providing visual, interactive maps of customer and account relationships and transaction flows. Link analysis now enables investigators to more readily identify patterns of activity that may involve extended rings of customers, counterparties, and transactions.
Moreover, insights from alert investigation, entity resolution, and link analysis are potentially consumable by an AI-driven "analyst" that would package the information into reports automatically created through natural language generation.
As these automated processes mature, and more AI is brought to bear upon analysis, alert decisions that are now made by humans will be within reach of the machines. In a very real sense, analyst decisions are already guided by automatically-generated risk scores, which can be used to prioritize alerts or even generate preliminary alert decisions. As the science behind risk scoring becomes more sophisticated and AI-driven, the closer we will get to computer-driven alert decisioning.
From the case management point of view, then, software robotics, machine learning, AI, and visualization can generate significant efficiencies in AML operations. Similar strides are being made in the application of advanced analytics and AI to model building and anomaly detection. While the technology will inevitably require much effort and practice to mature, it is not an overstatement to say that AML technology is already halfway to automation.