Although AI has become a popular topic, there has emerged a massive gap in the financial crime compliance. Numerous FCC teams cannot regulate the AI models that they implement. With the adoption of AI models in institutions, the issue of adequate monitoring of the models has become a matter of concern.
The research was conducted on 125 compliance and risk executives of international banks. It established that more than half of the technical issues that constrain AI growth in anti-financial crime initiatives deal with model governance. This highlights one of the greatest obstacles to operation by financial institutions.
The construction of an AI model is not the only step. The more sophisticated one is achieving some validation and operationalization of those models and sustaining them in the long run. Unluckily, most compliance teams do not have resources and governance structures that they require.
The greatest concern is limited or low-quality training data. Approximately, 91 percent of the interviewees put data quality as one of their five highest priority challenges. In the absence of clean and trustworthy data, AI models are prone to providing false alerts and increasing false-positives.
Another major hurdle is integration with the existing systems. Approximately 86 percent of the respondents claimed that they were struggling to associate AI models with the existing banking systems. Lack of proper integration slows down deployment, and requires compliance teams to do additional manual work.
Data Quality and Integration Challenges in AI Governance
It is common in many institutions that model outputs cannot be interpreted or trusted. This was considered to be a huge issue amongst the respondents (83%). Lack of explanations by compliance teams regarding why a model raised a flag on a transaction makes them unable to justify their actions to the regulators.
Explainability is now one of the important governance requirements. Black-box AI systems make it impossible to show adequate responsibility of the organization. The absence of transparency is a regulatory risk to financial institutions.
Almost three-quarters of the interviewees identified data and model governance as a distinct issue. Besides, 7 out of 10 reported a decline in model performance with time. Due to the rapid development of financial-crime strategies, constant monitoring and retraining is required.
A model that has been extensive in deployment can fail in the future. The old models may present compliance risks without frequent updating and monitoring. To maintain high performance, there is the need to have continuous governance.
Governance Challenges After AI Models Go Live
The report has also looked at some of the challenges that come with the transition of AI models in a production environment after testing. Some of the early issues like data quality remain, but new issues of governance become apparent at scale. It becomes more difficult to manage models deployed.
Approximately 43 per cent of the respondents expressed anxieties regarding post deployment model updates. Updates used by data-science teams can be time-consuming and reactive, owing to their heavy workload. This sluggishness is subjecting institutions to the new financial-crime risks.
The 38 percent of the respondents claimed that it is hard to maintain governance in various models. The larger the model inventories are, the harder it is to keep record of documentation, version control, and audit trails.
In the mean time, 33 percent of the participants indicated that the problem of interpreting model outputs is still there. Teams will need to constantly test the recommendations of AI systems even after the deployment.
Read : Why Health Care in America Is So Expensive Today
Building Strong Model Governance for Financial Crime
The report presents three fundamental counterparts of model governance of FCC teams. To start with, keep track of all things during the model lifecycle, including the intended purpose and data sources up to performance measures and updates.
Second, establish a reputation in AI outputs. The teams should have an opportunity to justify and support model decisions in case of an interview with regulators or auditors. Eligible AI is quickly turning into a business norm.
Third, maintain the effectiveness of the models once deployed. Periodic retraining would enable models to keep up to date as patterns of financial-crime evolve. Applications such as the Analytics Studio by Hawk are useful in tackling these issues. They provide automatic documentation, clear explanation of decisions and such tools that allow compliance teams to retrain models independently.
