Joe McNamara
As institutions grapple with the promise and complexity of artificial intelligence in compliance operations, two distinct approaches have emerged: backend augmentation and agentic solutions. Understanding when and how to deploy each approach is becoming increasingly crucial for organizations looking to harness AI effectively.
The Segmented Reality of AI Implementation
Rather than a one-size-fits-all solution, AI implementation in financial compliance requires a nuanced, segmented approach. The path forward depends heavily on organizations' current AI journey, their specific compliance pressures, and the nature of their operations.
Some processes lend themselves well to expert service models combining machine and human knowledge. In contrast, others are better suited to machine-first operations - where AI handles the bulk of recommendations while human experts provide quality assurance and context.
The Context Challenge
One of the most significant factors in successful AI deployment is context. AI agents are not hard-coded software solutions but depend heavily on the context provided by domain experts. This context comes in various forms - taxonomies, prompt libraries, and specialized knowledge that guide the AI's understanding and decision-making processes. Similarly, domain expertise remains essential, even in machine-first approaches. The challenge lies in understanding what constitutes expected outputs from AI systems. Organizations that lack domain expertise to evaluate AI-generated results often find themselves in trouble, accepting outputs they cannot properly assess or validate.
The key to successful utilization starts and ultimately is determined by the professionals on the other side of the screen, reviewing the output through the lens of practical expectation and experience of completing the work first-hand.
Building in Skepticism
While AI/ML is not a new concept in the financial services space, the evolution from traditional handcrafted models to more sophisticated AI systems has introduced new challenges around repeatability and consistency. Unlike earlier systems, where every variable was explicitly defined and transparent, modern AI models have accelerated at a pace where monitoring and documenting each decision (billions of them!) is no longer feasible, as legacy processes once allowed. This inevitably leads to operational results that resemble "black boxes" rather than the defendable, documented decisions required during regulatory review. Needless to say, this presents significant challenges for highly-regulated industries.
The solution lies in codifying skepticism into AI processes. Human beings are naturally skeptical, especially when investigating transactions or evaluating documents. This skepticism—the instinct to look for errors and potential problems—must be systematically built into AI systems through input and output validation processes. Just as an analyst may not yet have the complete picture of what they're reviewing but still "knows" something's not right, this same innate skepticism must carry through to the usage of AI tools and systems.
Organizations need to develop "skepticism libraries" that check for consistency, proper referencing, and accuracy at every stage of the AI workflow. So far, there's only one verified source for 100% authentic skepticism, and it's not embedded in your AI model. However, it's not out of reach either—these skeptical minds are sitting on your teams as we speak.
Process-First Implementation Strategy
Organizations should start with process mapping rather than jumping directly to technology deployment when considering AI implementation. Key questions to address include:
- How will AI be used within existing workflows?
- Where in the process will AI provide the most value?
- Who will be using AI tools?
- How will users interact with AI outputs?
- What level of explanation and transparency is required?
Understanding these process elements is crucial because while foundational models have made AI more accessible and reduced development time, the integration challenge remains a significant obstacle for institutions that have not taken the appropriate measures to scope and answer questions like the ones above.
The 70-30 Rule
A helpful framework for AI implementation suggests that technology represents only about 30% of the solution. The remaining 70% consists of people and processes. This perspective emphasizes that successful AI deployment isn't primarily a technology challenge but rather an organizational transformation that requires careful attention to human factors and workflow integration. The goal is not to remove the human from the equation, but rather to positively improve their position within it, and augment their critical thinking skills and healthy skepticism to make decision-making faster, more accurate, and repeatable over time. This can help build essential trust between AI, your people, and your regulator, while reinforcing the models’ abilities to learn and progress.
Building trust in AI systems requires a methodical approach through focused pilot programs. These pilots should be designed to test technical capabilities and validate the entire process of human-AI collaboration, ensuring that outputs are reliable, explainable, and actionable.
Looking Forward
As financial institutions continue to explore AI applications in compliance, the most successful implementations will likely be those that recognize AI as a powerful tool requiring expert guidance rather than replacing human expertise. The future lies not in choosing between human and machine capabilities but in thoughtfully orchestrating collaboration.
Organizations that invest time in understanding their processes, building appropriate skepticism into their systems, and maintaining strong domain expertise will be best positioned to realize AI's potential while managing its risks effectively.
The path forward requires patience, careful planning, and a commitment to getting the fundamentals right - starting with process, focusing on trust, and never losing sight of the critical role that human expertise plays in guiding and validating AI-powered solutions.
AI in the financial compliance space? Trust, but confirm.