There’s a new subset of artificial intelligence/machine learning (ML) that are taking over news feeds called Generative AI aka Gen AI. As compared to where traditional ML model output predicts, classify or cluster, Gen AI as the name suggests aims to create new content like text, video, audio, image, or code from the training data.

Encouraged by regulators around the world including FinCEN in the US, Financial Conduct Authority in the UK etc financial crimes compliance industry has been in forefront to adapt innovative technologies to catch bad actors and organized criminal network. In a recent press release European Union said - “Generative AI systems and virtual worlds are disruptive technologies with great potential.” (source).

Potential Applications

The success and maturity of financial crimes compliance programs across the globe has been a journey. Even after decades of public-private partnership, there are areas of amendment opportunities whether it is client risk assessment, suspicious transactions monitoring, case investigation, documentation regular reporting, and model management.

Below are some potential areas of application for GenAI in financial crime compliance.

1. Entity Risk Profile – Firms leverage multiple systems for KYC, sanctions, AML transaction monitoring etc which provides incomplete and ineffective entity profiles during case triaging and investigation. Thus, without full entity risk profile information including historical activities, suspicious behaviour, the case decisions are inefficient. Leveraging text generation area of GenAI called Large Language Model (LLM) meaningful and comprehensive entity profiles can be presented to the users for informed decisioning.

1. Case Narrative – Case documentation and narrative with or without regulatory reporting is generally completed by end users which takes longer, prone to human errors and is unproductive in providing key case details to law enforcements. Like entity profile LLM can be leveraged to generate meaningful case or suspicious activity reporting narrative which includes comprehensive information on triggered red-flags, investigation notes, research information, related cases/entities etc. 

1. Code Interpreters for Developers – Traditionally model development is a coding exercise followed by model review and operationalization. For large organizations end to end model development can more than 12 months. Gen AI can be used to generate code templates, frameworks, and libraries, providing developers with a head start in their projects.

1. Automating Model Documentation – Model management & documentation is fundamental to a robust FCC program and require data scientists and technical writers to document model code, input/out parameters, and full calibration processes. This can take days and can is prone to human error thus, an area which GenAI can help automate making the model management & governance much more efficient.

1. Dynamic Dashboards for Compliance Officers – Generally, reporting tools or backend queries are leveraged to research critical information about an entity or network specially during level 2 and special investigations by compliance officers. Gen AI suggests different data analyses that can be created. Compliance officers can also prompt their dashboard requirement in natural language, such as English. It also auto-suggests a suitable visualization based on the selected analysis.

There are other areas where GenAI or LLLM specifically could be applied though, above are some key once with possible highest impact.

Not a Magic Wand – Key Considerations   

Below limitations and a robust plan for manage the same should be factored part of the initiative.

1. Limited Talent – Gen AI is new thus, there is limited skill set and talent especially when you combine it with FCC domain expertise.

1. Risk of Hallucination - LLMs are prone to "hallucinations" - generating fictitious information, presented as factual or accurate.  Thus, output can be hard to control sometimes, and model might generate something which might not be what we are looking for thus, it is key to have robust model management.

1. Data Quality - Like any other technological initiative, Gen AI can only be as good as data thus, good data is key for quality of the generated content.

1. Quantity of Data - Furthermore, Gen AI leverages millions of features thus, requires a load of data for training.

1. 5.         High Compute Capacity – Since Gen AI requires loads of data for model training, the program would require very high compute capacity from infrastructure point of view. 

In addition to some of the known limitations outlined above, Gen AI may be prone to problems yet to be discovered or not fully understood.

How to Get Started?

Considering the freshness of GenAI the first step is to pick the right technology & business partners for the initiatives. A technology partner that brings both domain experience, data scientist skills and necessary infrastructure to manage the data, train, and deploy these models. Financial crime compliance domain expertise is key to ensure institution-specific business needs can be translated efficiently from the technology standpoint.

Furthermore, the right partners it is crucial to map a short and long-term roadmap for the adoption. Short-term roadmap should align with institutions’ specific business requirements and gaps, and success criteria to assess to effectiveness of those use cases.

For timely engagement from IT, business, model management and other stakeholders, senior management buy-in and investment should be agreed upon. Regulators are key stakeholders especially for complex initiatives like these once thus, keeping them loop and updating them about the roadmap, risks, and plan to mitigate those risks is must.

Finally, due to high data and compute capacity requirements, the technology needs for this initiative are expected extensive, driving higher costs. Therefore, aligning business value and benefits with investments with proper timelines of return on investments (ROI) is crucial for the scalability of the initiative enterprise wide.

Conclusion

It is estimated that, Gen AI could enable automation of up to 70 percent of business activities, across almost all occupations, between now and 2030 (source). Considering it is an evolving area the knowns limitations and risks should be well factored. Good news is that financial crimes compliance industry is be starting from scratch as there’s quite a lot of learning from AI/ML work done so far. As the interest grows the investment will rise in the technology, skill set etc. Partnerships between technology vendors, institutions, consulting firms, and regulators are key to leveraging the value of Gen AI.