How to address human bias when building out automated AML programs

In a previous post we wrote about the value of automating UBO identification as part of regulated entities managing their AML and KYC / DD processes.

Compliance departments at regulated entities need to keep these automated systems updated to match the quickly changing behaviour of bad actors and changes in regulations. 

Inability to effectively update automated systems can result in an increase of compliant customers being rejected or given a poor on-boarding experience (false positives) or low detection rates of bad actors. Invariably the failure to keep automated systems updated results in more manual intervention and higher costs.

But how does a Chief Risk and Compliance Officer (CRCO) ensure that no human biases are introduced into their automated review and workflow processes every time an automated system is adjusted?

Automated systems, machine learning and bias

Moreover more and more automated systems are beginning to use machine learning models (MLMs). These models are branded as un-biased however are programmed and set up by humans. How does a CRCO ensure that her MLMs are not impacted by human bias?

Ensuring that automated systems, whether legacy or MLM do not produce biased results ensure the regulated entity continues to operate ethically and ensures continuing customer trust and brand equity.

Entry of bias into automated systems

Most automated systems are dependent on the input of accurate and complete data to process workflow and make ongoing decisions. Where MLM is involved the systems need to be trained on a set of data. 

Data entry

The first entry of bias in automated systems is therefore in the incorrect or incomplete entry of data.

This can happen for a number of reasons including un-availability of complete or accurate data,  time restrictions and need for speed in setting up systems quickly and timing inconsistencies where systems are set up and launched operationally because of operational deadlines before a complete set of data has been collected, cleaned and prepared for input.

Entry of non representative data adds to the above risks of introducing bias. 

Human configuration of automated systems

Besides quality, timing and completeness of data bias can also occur through human configuration of automated systems. For example, while it is widely accepted that it is unethical to profile and configure for gender or ethnicity use of economic data points such as employment, address and total assets can inadvertently introduce bias into systems that result in false positives, poor on-boarding experience or the poor filtering of bad actors.

Bias in triage and action triggers

Human bias can also impact the automated or manual action taken when one or more criteria have been triggered. The CRCO must ultimately decide which triggered alerts are investigated further and which to report to the regulator. It’s important during these triage moments to ensure no human cognitive biases such as availability bias, prospect theory or anchoring impact the triage decision.

Reducing bias in automated AML systems

Approaching and tackling the topic of bias in automated AML systems requires close collaboration between the compliance and data science teams within regulated entities. 

Importance of compliance and data science collaboration

Firstly, data science teams need to have a clear understanding from compliance colleagues on the regulatory requirements that data models need to comply with and also generally the values and principles with which the compliance department operates.

Importance of audit and governance

The development and configuration of the automated system and the quality and completeness of the input data should be transparent, visible. These activities should also be auditable and come under the regulated entities governance processes to ensure no individual or group bias is allowed to creep in.

White box vs black box models

Where possible regulated entities should attempt to build only interpretable or white box models and avoid black box models. White box models while more effort in setting up provide a better understanding and audit trail in tracking the possible introduction of biases. Where black box models are unavoidable these should be documented extensively and in a manner that is easy to understand, action and report to the regulator if needed.

Continuous monitoring and testing

Automated systems and data models need to be continuously monitored and re-assessed for suitability and matched against the original criteria at the time of their creation. Data and the relationship between the data in the systems changes with time especially with the introduction of new products, services and expansion into new geographies. These changes can cause an automated system to start becoming more inefficient or introducing unwanted biases. 

Automated system output should be periodically checked with other data points to ensure no bias has crept in. For example model outputs can be cross checked against data such as race, address, religion or income. 

Want a free consultation on how to avoid bias in your automated AML systems? Contact us or a no obligation chat.