In financial services, regulation isn’t just a rulebook: it’s a rule maker.
Regulatory compliance can be overwhelming, with financial institutions currently facing unprecedented volumes of regulatory documents to monitor and implement across the globe. Managing this complexity manually is no longer feasible. The solution to navigating compliance frameworks is by adopting artificial intelligence (AI).
AI is playing an increasingly crucial role in supporting compliance teams, for instance, by processing vast amounts of regulatory data, monitoring transactions and automating routine but essential compliance tasks. More specifically, it enables institutions to remain compliant with evolving regulations, while being efficient with limited resources. As not all AI models are equally suited for high-stakes applications, such as regulatory compliance, the key to the AI-led efficiency is using a highly-targeted layered approach, with surgical precision applying different kinds of AI models for different purposes.
Battle of the models
Using AI without a clear understanding of its specific capabilities and limitations is akin to prescribing treatment without a diagnosis. In the same way a doctor would not administer precise chemotherapy for a simple infection, but would rather use broad-spectrum antibiotics, and vice versa, organisations must first identify desired outcomes, and then discern the unique attributes of AI models to apply them effectively. It’s imperative to identify where these models offer distinct efficiencies and to integrate them thoughtfully, ensuring they enhance operations without introducing unforeseen risks.
Large Language Models (LLMs):
Large Language Models have billions of parameters, making the models brilliant at understanding context, nuances and complex language patterns. These models are trained on vast datasets and can demonstrate impressive capabilities in generating human-like text while understanding subtle implications in regulatory documents. Once trained, they can be refined for specific tasks, such as creating systems that recognise patterns across enormous amounts of text corpora.
Specifically for compliance functions, LLMs excel in regulatory change management. The models can analyse new regulations, identify the potential impact on the organisation and generate roadmaps for implementation. The ability of LLMs to understand semantic meaning makes them valuable for compliance risk assessments and policy interpretations, which in turn helps compliance teams understand how regulatory changes can affect the business.
However, LLMs have limitations for financial institutions. For one, they are expensive. They also have substantial computational requirements, which makes maintenance difficult, especially if they are used for real-time applications where immediate responses are key. Privacy can also present an issue if a model requires sensitive data to run third-party APIs.
However, the most concerning factor for compliance teams is that there are challenges in explainability and maintaining an audit trail, which are critical requirements in regulated environments. After all, if teams are not clear about how a conclusion is reached by AI but rely on it anyway, this will spell trouble when mistakes are made.
Small Language Models (SLMs):
Small Language Models provide a completely different application to AI in compliance. Namely, these models prioritise efficiency and specialisation over breadth, particularly given their lower parameter count, which rests in the millions, rather than the billions seen in LLMs. The first difference this creates is that SLMs are designed for focused tasks and can be trained on an organisation’s own domain-specific data.
As a result, SLM mechanics involve more targeted learning on smaller datasets, in turn producing more predictable and consistent outputs for specific compliance functions. Specifically, SLMs find their niche in transaction monitoring and real-time screening of financial activities. SLMs also excel at pattern detection in structured data, making them ideal for spotting anomalies that may indicate financial crimes. Additionally, their smaller data footprint also alleviates some of the privacy concerns associated with larger models.
However, SLMs come with considerable limitations in their scope and flexibility. They often lack the contextual understanding of larger models, and as a result, struggle with specific scenarios that may not be represented in training data. The effectiveness of SLMs depends wholly on the quality and comprehensiveness of the training data sets, making them potentially less reliable when facing unexpected regulatory changes or unusual compliance scenarios.
Creating synergy with multiple models:
An effective AI compliance strategy must employ a layered approach that leverages both LLMs and SLMs in complementary roles. Having this architecture allows financial institutions to match the right model to each task based on its specific requirements for compliance.
For example, an SLM can perform the initial screening of transactions for potential issues and flag the most concerning cases, then the LLM can take over to perform a deeper analysis on the information provided.
Layering your approach results in a more efficient operation. The ideal balance is to reserve LLMs for complex tasks that require deep contextual understanding and deploy SLMs for high-volume, routine compliance activities. This way, institutions can optimise performance and cost. According to reports from the World Economic Forum, financial institutions that implement combined AI strategies see efficiency improvements of 35-40% in compliance operations.
Inevitably, there are challenges with integration. It is crucial to ensure these models can work together seamlessly within existing compliance frameworks. Successful implementation requires a carefully designed system that allows the models to share information effectively, while maintaining appropriate checks and balances. Financial institutions must develop clear protocols for when to use each type of model, and how to manage any potentially conflicting outputs.
How adaptable are these models when faced with regulatory change?
As we know, regulations adapt and evolve, and new requirements regularly emerge across different jurisdictions. Having multi-model approaches provides the crucial adaptability this environment requires, combining the quick responses of SLMs and the deep analytics of LLMs. This flexibility is invaluable when addressing cross-border regulatory requirements, where compliance teams need to navigate complex and sometimes contradictory regulations.
In this situation, many financial institutions have found that SLMs can be quickly retrained or adjusted to accommodate specific regulatory changes, while LLMs will provide the broader context necessary for comprehensive compliance strategies. A McKinsey study indicates that institutions that employ a dual approach respond 60% faster to regulatory changes than those relying on single-model solutions. This can help reduce the risk of enforcement fines, and saves potential reputational damage caused by falling foul of regulators.
Implementation strategies:
Integrating AI into a compliance function must be a considered process. From an operational standpoint, few compliance teams will have immediate access to the full computational power required to use multiple AI models. This is where it’s crucial to evaluate existing systems and identify the points where different AI models can be integrated. For some firms, this may be where a cloud solution could be helpful, as this could help increase access to computing power that is not available in-house, and scale it as needed.
Secondly, managing data for multi-model environments is crucial. Institutions need to develop clear protocols for how data flows between models to ensure accuracy and appropriate privacy controls. Data governance frameworks that maintain regulatory compliance must be established, while also enabling AI systems to access the information needed to operate efficiently. A good place to start is defining use cases where AI can deliver immediate value, such as suspicious activity monitoring, or tracking regulatory updates.
Needless to say, AI in financial compliance needs to include strong governance frameworks. To do this, compliance teams must establish clear lines of responsibility for AI systems, including model selection, validation and ongoing monitoring. Senior leadership and compliance officers must maintain appropriate oversight of AI-driven processes to ensure alignment with regulatory expectations and possible institutional risks.
Future directions:
What is the future of AI in financial compliance? All signs point to increasingly sophisticated model combinations. Hybrid models that combine the strengths of both large and small models provide the best of both worlds for compliance teams.
Attitudes towards AI in compliance are changing for the better, with many regulators now actively encouraging responsible AI adoption. The keyword here is “responsible”.
Looking ahead, the most successful financial institutions will be the ones that view AI not as a single solution but as a toolkit to increase efficiency, improve adaptability, and strengthen their compliance function as a whole.
