by Matthew Bazeley, Co-Founder
In June 2023 Google Cloud released a new product for detecting Anti-Money Laundering aimed at financial services customers. As the name suggests, AML AI is a machine learning solution in a problem space traditionally dominated by rules based systems.
The product was created and developed in partnership with HSBC over multiple years, and was launched to the wider industry following a successful private trial of the tool by HSBC.
Google Cloud's AML AI has significantly improved HSBC's AML detection capability. Google's models are already demonstrating the tremendous potential of machine learning to transform anti-financial crime efforts in the industry at large
For HSBC, AML AI identifies two to four times as much suspicious activity as the previous system, while reducing the number of alerts by 60% 1. This is not a theoretical backtest on a carefully curated set of data, this is a real, achieved benefit sustained over years.
Google’s AML AI is a ready-to-run model pipeline which allows organisations to build, test and deploy production models to detect money laundering. The product comes with a Google feature set for detecting AML risk and an engineering and analytical pipeline for creating and running AML models abstracted behind an API. Once a bank has their data on GCP they can call the AML AI API to automatically train, tune, backtest and predict their customer’s AML risk using Google’s pipelines.
Since the product is a pipeline for creating models rather than a model itself, it uses client data as input in model tuning and training to build AML AI models bespoke for each customer. Further, since AML AI uses a supervised learning approach, to use it a bank must have at least some customers labelled as AML risk (SARs or Exits) so the model can learn what AML activity ‘looks like’. Luckily for everyone, given the lack of comprehensive labelling in the financial crime space, tests have shown that bank’s don’t need many AML example customers to see good results, with broad risk coverage, from AML AI models.
Google’s models do not score each transaction for risk, instead AML AI considers customer activity holistically to produce an overall AML risk score for each customer using 2 years of data for risk identification and context. This allows it to better detect complex money laundering patterns which have always been out of reach of rules based systems.
AML AI requires the following data to train and use models2
Data needs to be on GCP and formatted into an AML AI dataset. Once a dataset has been created and registered, AML AI processes can be triggered.
AML AI has four different types of processing for different stages of the model development lifecycle. Each process is triggered and managed by API calls and all operations require an input AML AI dataset. A single AML AI dataset can be used for different types of operation, and for the same type of operation multiple times.
These operations are distinct, and it is not normally necessary or advisable to run through the entire flow every time. For example, an Engine Config can be used to train many models and a trained model can be used for many backtests or predictions.
| AML AI Process / Operation | Purpose | Inputs |
|---|---|---|
| Engine Configuration | Creates a configured version of the AML AI product with hyperparameters tuned against the bank’s data. This ‘EngineConfig’ can be used to train AML AI models. | 1. An AML AI dataset 2. Engine Config 3. Parameters & Objectives |
| Model Training | Creates a trained AML AI model which can be used to predict customer AML risk and for backtest operations | 1. An EngineConfig 2. An AML AI dataset 3. Training Parameters |
| Backtest | Tests the performance of a trained AML AI model against previous AML risk events (SARs and Exits). | 1. A trained AML AI model 2. An AML AI dataset 3. Backtest Parameters and Objectives |
| Prediction | Outputs an AML risk score for each customer using an AML AI model. Also outputs data explaining what feature groups contributed to the score, and metrics to support model monitoring in production. | 1. A trained AML AI model 2. An AML AI dataset 3. Prediction Parameters |
Google provides a suite of detailed documentation for model governance purposes to help banks understand their model and meet model governance requirements. Documentation includes information on Model Architecture, Feature Types, Labelling, Training, Tuning, Model Evaluation and Risk Typologies. The AML AI operations also output data to support the model governance process as part of their standard output.
AML AI offers banks a way to use AI to score their customers holistically for AML risk and provides well structured frameworks for them to interface with.
Our people at Groundtruth have extensive experience building, testing and deploying Google AML AI, and we have evaluated its performance countless times across regions and lines of business. Groundtruth are uniquely placed to best help our customers understand AML AI, integrate it into their systems and evaluate its performance against incumbent processes.
To learn more about the AML AI product and how it could be used at your organisation, contact us today.
GCP Technical Documentation - AML AI Creating and Evaluating Models
WSJ - Google Cloud Launches Anti-Money-Laundering Tool for Banks, Betting on the Power of AI
Fighting money launderers with artificial intelligence at HSBC
For decades traditional rule based Anti-Money Laundering products have offered potential compliance to AML regulations but ineffective solutions. AI is now offering financial institutions the opportunity to implement more effective and cheaper controls against financial crime while remaining compliant.
Read moreAlongside the operational and risk benefits of deploying a robust and effective AML model, Google’s extensive experience of operating machine learning and AI for decades has embedded MLops and devops capabilities which can be transformative for your ability to rapidly react to new and emergent risks.
Read more